GenAI catch-up research note 2026-09-27

🤖 Generated by Claude Opus 5.5

• Window: 2026-09-12 to 2026-09-27 (the first run, with no previous note, so a 14-day bootstrap window; 221 candidates).
• Sources that failed to fetch: none (all 38 fetched).
• This is the first run, so no theme continues from an earlier one.

Candidates and selection

• Selected (go into the report):
  1. Claude Opus 5.5 and GPT-6 Sol/Luna released on the same day, and the price war that followed (anchor 09-22, impact 2 / potential 2).
  2. TypeSafe's Jev (a decision model) released, with integrations, clones and field tests within two weeks (anchor 09-15, 2 / 2).
  3. Claude Code support for AGENTS.md, and the upkeep of instruction files (prompt-audit) (anchor 09-18, 2 / 2).
  4. Controls for unattended runs from four coding-agent vendors (anchor 09-19 to 09-25, 2 / 2).
  5. Controls in the agent platforms of Azure, Google and AWS (anchor 09-15 to 09-24, 2 / 2). - 1 and 2 tie; 1 ranks higher because the new default model and prices reach the users of every tool at once.
• Dropped (input for the next run's judgement):
  6. Existing REST APIs and SaaS knowledge turned into MCP tools and Skills (anchor 09-23 to 09-25, 1 / 2) — each feature is small, and none reaches the top five.
  7. The debate over "development without hand-written code" (anchor 09-24, 2 / 1) — the debate is broad, but as a change in practice it is one company's policy.
  8. Practical know-how on harness design, from Japan (anchor 09-16, 1 / 2) — the in-window anchors are case studies and talk slides, not a primary change of specification.
  9. A big release week for LangSmith (anchor 09-24, 1 / 1) — a platform update from one company.
  10. Operational changes to the Claude Platform API (anchor 09-14 to 09-24, 1 / 1) — no more than a supplement to theme 1.
  11. A run of Kiro updates (anchor 09-14 to 09-24, 1 / 1) — apart from 2.24, which theme 4 covers, improvements to one product.
  12. A batch of enterprise updates to GitHub Copilot (anchor 09-23 to 09-25, 1 / 1) — apart from the two items in theme 4, administrative features.
  13. Google model and platform updates (anchor 09-15 to 09-24, 1 / 1) — centred on voice models, away from the core of this run's tracks.
  14. Open-weight models and inference efficiency (anchor 09-14 to 09-23, 1 / 1) — a thin link to everyday practice in Japan.
  15. Claude Code cloud sessions made official (anchor 09-24, 1 / 1) — a rename and a tidy-up of settings.
  16. Codex model and CLI updates (anchor 09-14 to 09-26, 1 / 1) — mostly deprecation notices that affect few readers.
  17. Practical material on evals and benchmarks (anchor 09-16 to 09-18, 1 / 2) — single documents whose momentum could not be confirmed.
  18. Google ADK and Antigravity updates (undated, 1 / 1) — the feed carries no dates, so no anchor can be fixed.
  19. Other one-off items (1 / 0 to 1) — listed below.
  20. Out of scope — listed below.

1. 🚀 Claude Opus 5.5 and GPT-6 Sol/Luna shipped on the same day, starting an API price war

• Selected / impact 2 / potential 2 / anchor 09-22 / track: both.
• Feed items in this cluster:
  • claude-code-changelog 09-22 2.1.280
  • claude-platform-release-notes 09-22 September 22, 2026
  • openai-news 09-22 Introducing GPT-6 Sol and Luna
  • openai-news 09-22 Better prompt caching for GPT-6
  • openai-chatgpt-codex-changelog 09-22 GPT-6 Sol and Luna in Codex and ChatGPT Work
  • aws-whats-new 09-22 OpenAI GPT-6 Sol and GPT-6 Luna are now generally available on Amazon Bedrock
  • aws-ml-blog-bedrock 09-22 Claude Opus 5.5 is now available on AWS
  • gcp-agent-platform-release-notes 09-22 September 22, 2026
  • simon-willison 09-22 Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war
  • ainews 09-23 Claude Opus 5.5, the new default model for AINews
  • classmethod-claude-code 09-23 Claude Code v2.1.280 の主要アップデート (key updates in Claude Code v2.1.280)
  • hatena-llm 09-23 What a task costs on Opus 5.5
  • zenn-trending 09-24 王の凱旋—Opus 5.5 React習熟度ベンチマーク— (the king's triumphant return: an Opus 5.5 React proficiency benchmark)
  • zenn-trending 09-23 Claude Opus 5.5によるピクセルアニメーション生成の検証 (testing pixel-animation generation with Claude Opus 5.5)
  • github-changelog-copilot 09-25 GitHub Copilot weekly releases — September 21

Collected information

[A] Anthropic — Claude Opus 5.5 (2026-09-22)

https://www.anthropic.com/news/claude-opus-5-5

• On typical workloads it "costs 40% less to run than Opus 5", and it generates output "more than 30% faster".
• Terminal-Bench 4.0: 66.4% (Opus 5: 52.3%).
• FrontierCode v1.1 (Main): 54.4% (Opus 5: 48.0%).
• CursorBench 4.0: 57.8% (Opus 5: 46.6%).
• GDPval-AA v2.1: 1846 Elo (Opus 5: 1708).
• OSWorld 2.0: 81.8%, ahead of Fable 5.1 at 80.7%.
• Anthropic positions it as performing "at the level of Claude Fable 5.1 on most work".
• Pricing per million tokens: input $4 / output $20 (Opus 5: $5 / $25), cache reads $0.20 ($0.50), cache writes $5 ($6.25), and Fast mode $8 / $40.
• Available on the Claude API, AWS, Google Cloud, Microsoft Azure and Claude Code, and on the Pro, Max, Team and Enterprise plans.
• Safety: the best scores of any model to date on the automated behavioural audit, "a classifier that screens every action before it runs", attempts to circumvent boundaries "around 85% less often than Opus 5", biology and cybersecurity safeguards that "fall back to another model transparently" (most cybersecurity tasks are "re-routed to Opus 4.8"), and "preserved thinking, the anti-distillation safeguard".

[A] Anthropic — Claude Platform release notes, September 22, 2026 (2026-09-22)

https://platform.claude.com/docs/en/release-notes/overview#september-22-2026

• Claude Opus 5.5 (claude-opus-5-5) is "a model for long-running agentic coding and knowledge work".
• It has a "1M token context window by default, 128k max output tokens, and always-on adaptive thinking".
• Pricing is $4 / $20 per MTok (Claude Opus 5 is $5 / $25).

[A] Anthropic — Claude Code changelog 2.1.280 (2026-09-22)

https://code.claude.com/docs/en/changelog#2-1-280

• Opus 5.5 is "now the default Opus model" (1M context, $4 / $20 per Mtok, $0.20 per Mtok for cache reads).
• The default model on Pro and Team Standard changed from Sonnet to Opus, "matching Max, Team Premium, and Enterprise".
• The deprecated TaskOutput tool was removed (Claude uses Read instead).
• Enter and Esc now accept and cancel dialogs, and y / n are no longer bound by default.
• Opus 4.7, Opus 4.8 and Fable 5 "stop holding their launch-default effort over" an explicit /effort setting.
• The new CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH changes the cap on the length of MCP tool descriptions.

[A] OpenAI — Introducing GPT-6 Sol and Luna (2026-09-22)

https://openai.com/index/introducing-gpt-6-sol-and-luna

• Sol and Luna were trained with methods similar to Astra's and are positioned to push the cost-efficiency frontier forward.
• API prices are 50% below GPT-5.6's promotional pricing: Sol $4 → $2 / $20 → $10, Luna $0.20 → $0.10 / $1.20 → $0.50.
• On AutomationBench (47 tools; business workflows in sales, marketing, operations, support, finance and HR), Sol (xhigh) scores 33.2% at $0.27 per task, Astra (low) 30.3% at 3.9 times Sol's cost, Opus 5 (max) 26.9% at 11.1 times, and Fable 5.1 (max, with Opus 5 fallback) 31.4% at more than 8.9 times (the fallback cost is not counted).
• On Agents' Last Exam V1, Sol (max) scores 56.4%, beating Opus 5's best score at 60% lower cost.
• On internal factuality evals, Sol makes about half as many errors as the previous generation, and Luna at high effort matches GPT-5.6 Sol at about a hundredth of the cost.
• On FrontierCode 1.1 Main, Sol improves substantially on GPT-5.6 Sol and matches Fable 5.1 (xhigh) at a far lower cost.
• On DeepSWE v1.1, Sol (max) scores 68.8%, 1.1 points behind Fable 5 (xhigh) at 69.9%, for about 80% less cost.
• On DeepSWE, Luna (max) scores 66.6%, on a par with Opus 5 and Fable 5 (medium), at 93% less cost than Opus 5 and 96% less than Fable 5.
• On OSWorld 2.0 offline, Sol (xhigh) scores 60.5% against 60.3% for Opus 5 (medium), for about 80% less cost.
• Internal use: at API prices, the median researcher spends $600 a day and the 90th percentile $7,000.
• Caching: higher hit rates by default, cache reads at a 90% discount, a Prompt Caching Dashboard and diagnostic tools, cached context kept when reasoning effort or tool enablement changes, and explicit breakpoints.
• GitHub cut the share of prompt tokens that need fresh processing by more than 50% over several months.
• Availability: ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu from launch day, Luna for Free and Go in the desktop app, and the API as gpt-6-sol / gpt-6-luna.

[A] AWS — Claude Opus 5.5 is now available on AWS (2026-09-22)

https://aws.amazon.com/blogs/machine-learning/claude-opus-5-5-is-now-available-on-aws/

• The Bedrock model ID is global.anthropic.claude-opus-5-5.
• Regions: US East (N. Virginia) and AP Southeast (Melbourne) on bedrock-mantle, plus the US, EU, AU and JP geographic cross-Region inference profiles on bedrock-runtime.
• "Adaptive thinking is always on, and Opus 5.5 decides how much reasoning each task needs."
• It is "the first Opus model that comes with safety classifiers" for "biology, cyber security, and AI development".
• It is also available "through Claude Platform on AWS in North America".

[A] AWS — Bring more intelligence to everyday work with GPT-6 Sol and GPT-6 Luna on Amazon Bedrock (2026-09-22)

https://aws.amazon.com/blogs/machine-learning/bring-more-intelligence-to-everyday-work-with-gpt-6-sol-and-gpt-6-luna-on-amazon-bedrock/

• Astra is for "the most ambitious projects, where achieving the highest-quality result matters more than cost", Sol for "demanding tasks that recur throughout development and operations", and Luna for "focused, repeatable tasks" at "high volume".
• Sol made "approximately half as many factual mistakes as GPT-5.6 Sol".
• On Bedrock, Sol and Luna "support explicit prompt caching", and reasoning effort can be adjusted "per request".
• VPC endpoints, IAM, CloudTrail, and the option of zero data retention.

[A] Google Cloud — Gemini Enterprise Agent Platform release notes, September 22, 2026 (2026-09-22)

https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes

• Claude Opus 5.5 is available in Model Garden.

[B] Simon Willison — Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war (2026-09-22)

https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna/

• Anthropic released Opus 5.5 first, and OpenAI released Sol and Luna "around an hour later".
• Prices (input / cached input / output): Luna $0.10 / $0.01 / $0.50, Sol $2 / $0.20 / $10, Opus 5.5 $4 / $0.20 / $20.
• Opus 5.5 is 20% cheaper on input, and "the price for cache reads fell 60%".
• Luna is "one of the cheapest models OpenAI have ever released", beaten only by GPT-4.1 Nano and GPT-5 Nano.
• Opus 5.5 has "a 128,000 maximum output token limit".
• Opus 5.5 is "cheaper per token than Opus 5.0 with the intelligence of Fable 5.1" and "looks like it addresses the biggest complaints people had about Opus in terms of its communication style".
• At "max" thinking, though, it "over-thinks to the point of breaking": on an SVG prompt it used up its output tokens while still reasoning and "failed to return a response".
• At "one tenth" of the price of Haiku 4.5 ($1 / $5), Luna pushes the price competition across tiers.
• GPT-5.6 has "a scheduled 25% price increase for November".

[B] Latent Space (AINews) — Claude Opus 5.5, the new default model for AINews — and everybody cuts prices 40-50% (2026-09-23)

https://www.latent.space/p/ainews-claude-opus-55-the-new-default

• Opus 5.5 "performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5", runs about 30% faster, and is about 40% cheaper per task at default effort.
• It beats Fable 5.1 on most benchmarks and is "stronger than Astra on CursorBench, KWBench, and OSWorld".
• Vals Index #1 (up two places from Opus 5), FrontierSWE 62.3% (#2 behind Astra's 65.5%), FrontierCode 1.1 65.3% on Extended (taking #1 from Fable 5), CursorBench 57.8% (Max), and ParseBench 93.9% (+7 points over Opus 5).
• Session limits are up 20%, and subscribers get "a banked rate-limit reset they can use whenever they choose".
• AINews switched to Opus 5.5 immediately, finding that "the difference is night and day" (it "puts the most important information up front and follows the writing rules you give it").
• OpenAI's new models launched "50% lower than GPT-5.6" against a 20% cut for Opus 5.5, and the "40% cheaper" claim "applies to default (medium) settings" (at max effort, higher token use offsets it: "higher token usage alone would raise cost per task about 80%").
• An effort-scaling anomaly was reported: "on an agentic coding chart, xhigh effort costs about 2.8x more than medium for a 3.2-point lower score".
• Reactions split along the lines of "Anthropic led on capability surprise, OpenAI on price", and some questioned their trust in Opus releases (one user "no longer trusts Opus releases to feel better").

[B] Addy Osmani (claude.dev) — What a task costs on Opus 5.5 (2026-09-25)

https://claude.dev/blog/what-a-task-costs-on-opus-5-5/

• "Two models with the same cost per token can cost very different amounts on the same task", so cost is measured per task.
• Worked example: a task that sends about 2.8M input tokens over 40 turns (about 70K per turn on average) costs about $1.62 in input, and the same task in 25 turns about $1.02.
• If high effort adds 20K thinking tokens across a task, that is $0.40 on Opus 5.5, about the same as "a retry loop of ten turns at 100K of cached context".
• In an internal migration test ("an internal customer support benchmark of 44 tickets"), moving from Opus 4.8 to Opus 5.5 cut cost by about 18%, and running prompt-audit cut it "by a further 9%".
• With the same token counts on both models (the price change alone), the session "costs about 31% less" on Opus 5.5.
• Turns drive cost: a turn "costs more than the tokens it adds, because it resends everything before it", so "the cheapest turn is the one you don't need".
• It cautions that "the figures in this post are illustrations" and that "your own numbers are the ones to trust".

[B] Classmethod — Key updates in Claude Code v2.1.280 (2026-09-23, in Japanese)

https://dev.classmethod.jp/articles/20260923-cc-updates-v2-1-280/

• The default model on Pro and Team Standard changed from Sonnet to Opus, and Opus now costs 20% less.
• Dialogs are now confirmed and cancelled with Enter and Esc, and an effortLevel saved in the old format no longer applies to new models.
• Opus 5.5 was used both with --model opus and with no model specified.
• Because the modelUsage keys differ between the two cases, aggregating logs and costs by canonicalModel is the safer choice.

[B] uhyo — The king's triumphant return: an Opus 5.5 React proficiency benchmark (2026-09-25, in Japanese)

https://zenn.dev/uhyo/articles/react-profession-bench-17

• Ranking: Opus 5.5 max 94.46, Opus 5.5 high 90.79, Opus 5 max 90.28, Opus 5 high 87.00, Fable 5 max 86.67, Fable 5.1 high 86.44.
• Opus 5.5 beats the previous generation's max, the previous record, and at max it pulls more than 4 points clear of it.
• Its scores are high but not as reproducible as Opus 5's.
• The evaluation covers 13 specs and 78 implementations, the grader is Claude Sonnet 4.6, and the implementation timeout went from 45 to 90 minutes.
• The TypeScript quality criterion is saturated at full marks and needs a redesign in v2.
• The effort curve has flattened: high is fine for everyday use, and max is an option for buying extra points.

[C] Zenn (feed) — Articles testing Opus 5.5's generative abilities (2026-09-23, in Japanese)

• Two articles trended on the same day: one testing pixel-animation generation, and one that gave Opus 5.5 a prompt written for video generation and had it make an explanatory animation (supplementary).

Notes

• The FrontierCode v1.1 figures of Anthropic (54.4%) and AINews (65.3%) come from different sets: Anthropic reports the Main set, AINews (citing Cognition) the Extended set.
• The "40% cheaper" figure is per task at default effort; the per-token price cut is 20%.
• OpenAI's benchmarks are mostly its own evaluations, and OpenAI itself notes that the cost of Fable 5.1's fallback is not counted.
• Simon Willison and AINews both report problems with Opus 5.5 at max effort, and the information in hand supports running it at high by default.
• According to the feed, the Copilot weekly release (week of 09-21) also added Opus 5.5, but the article itself was not fetched.

Ideas for practice and output

• Run Opus 5, Opus 5.5 and GPT-6 Sol at default effort on your own eval set (10 to 20 cases), and put turns, output tokens and cost per task into a single table.
• Call global.anthropic.claude-opus-5-5 once through Bedrock's JP cross-Region inference and check how the model ID and billing show up.
• Shortlist Luna for high-volume classification and extraction, put a cache breakpoint on the shared prefix, and watch the hit rate on the dashboard.
• Move Claude Code cost reporting onto canonicalModel, pin the default effort at high, and measure the difference from max.

2. 🎯 TypeSafe released Jev, a decision model that returns typed judgements with calibrated probabilities instead of text, and integrations, clones and field tests followed within two weeks

• Selected / impact 2 / potential 2 / anchor 09-15 / track: GenAI engineering.
• Feed items in this cluster:
  • simon-willison 09-21 Jev introduces a new shape of LLM - System One, aka Decision Models
  • ainews 09-19 Here are 6 Clones of Jev in 2 days
  • langchain-blog 09-23 What Is Jev? A Guide to TypeSafe AI's System One Model
  • langchain-blog 09-25 Building Production Agents with Jev and LangGraph
  • laiso-newsletter 09-17 #21 Jev
  • laiso-newsletter 09-25 System One Adapterを使ったJevとローカルモデルの分類タスク比較 (comparing Jev and local models on classification tasks with System One Adapter)
  • hatena-llm 09-24 みんなJevの話してる。やってないの俺だけ (everyone is talking about Jev and I'm the only one who hasn't tried it)
  • hatena-llm 09-25 Ollaya · Run decision models locally
  • zenn-trending 09-25 文章を返さないAI「Jev」を試してみた (trying Jev, an AI that returns no prose)
  • zenn-trending 09-25 Jevにゲームをやらせるな！LLMを蒸留してみよう！ (don't make Jev play games, try distilling an LLM instead)
  • hn-llm 09-17 LLM Classification Is Feature Engineering

Collected information

[A] TypeSafe AI — Introducing System One Models & Jev (2026-09-15)

https://typesafe.ai/blog/introducing-system-one-models-and-jev

• System One models are "a new class of frontier models built to make fast, structured decisions that software can use directly".
• Jev follows the pattern "unstructured state in, typed probabilistic decisions out", and it samples in parallel, generating "all outputs in a single query" instead of one token at a time.
• Because the "possible outputs and structure are defined in advance", the model "can't hallucinate" and "never makes type errors" by construction.
• Training uses RLCD (Reinforcement Learning for Calibrated Decisions) instead of RLHF / RLVR, aiming for calibration where "higher confidence means higher accuracy".
• On System One tasks Jev is "40x-200x faster" than frontier LLMs, answering end to end in "70ms-500ms" against "3 to 329 seconds" for frontier models.
• Input costs "$0.042 / MTok" and output is free, "too cheap to meter", while frontier models charge "from $0.20 to $10 / MTok" for input.
• Early access brings developers "off the waitlist" step by step, and Jev is used from console.typesafe.ai.
• Because Jev "gives up string generation", it cannot produce free-form text such as chat replies or generated code.

[A] Vercel — TypeSafe AI's Jev now available on AI Gateway (2026-09-16)

https://vercel.com/changelog/typesafe-ai-jev-now-available-on-ai-gateway

• The model ID is typesafe-ai/jev.
• AI SDK 7 "exposes Jev through the experimental evaluate API" (experimental_evaluate), supported from 7.0.105 onwards.
• Question types: "Choice selects an option, Score grades an ordered rubric, and Boolean estimates the probability of true".
• TypeSafe reports that Jev "was up to 193.6x faster and 444.6x cheaper than LLMs on its workflow evaluations".

[A] Vercel — AI Gateway now supports TypeSafe clients and an HTTP API for Jev (2026-09-21)

https://vercel.com/changelog/ai-gateway-now-supports-typesafe-clients-and-http-api-for-jev

• TypeSafe-compatible API: existing TypeSafe clients only "change the base URL and API key", and "your systemOne calls, noul questions, and response shapes stay exactly as they are".
• HTTP API: a POST to /v1/evaluate passes the model ID, the state and the questions (boolean / choice / score).
• On all three paths, requests "are billed through AI Gateway", so "they appear alongside your other model calls in usage and observability".

[B] LangChain — What Is Jev? A Guide to TypeSafe AI's System One Model (2026-09-17)

https://www.langchain.com/blog/building-a-harness-with-jev

• Three places where Jev stands in for an LLM call: model routing (a "fast" or a "powerful" model, chosen by how complex the request is), tool-call checks (AutoModeMiddleware blocks risky tool calls "before the tool executes"), and classification.
• The split is to "use an LLM for open-ended reasoning and generation, and Jev for fast, structured decisions along the way".
• TypeSafeClassifier from langchain-typesafe is called as classifier.invoke({"state": ..., "questions": {"urgent": Noul(instructions=...)}}).
• Question types are Choice, Score and Noul (yes or no), and several questions are evaluated in parallel in one request.
• TypeSafe reports "up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks".
• Jev "doesn't generate text" and "isn't a drop-in replacement for an LLM".

[B] LangChain — Building Production Agents with Jev and LangGraph (2026-09-25)

https://www.langchain.com/blog/building-prod-with-jev-and-langgraph

• Jev sits in LangGraph as a classification and routing node: "Code keeps the structure and handles exact computation, and a model sits only at the branches that need semantic judgment".
• In a document-review example, Jev answers three yes/no questions per page (is it responsive, does it contain personal information, might it be privileged), and the answers send the page to an LLM for redaction, to a human review queue or straight through.
• The rule is "cheap by default, frontier on exception": anything "below a 0.7 confidence threshold falls back to an LLM", and high-risk decisions pause for a human through LangGraph interrupts.
• Jev was "5–6x faster on the classification step", and Stagehand's median act() latency "dropped from 1.97 seconds to 0.46 seconds, about 4.3x faster".
• Across 100 repeated runs Jev's scores "barely moved", "far less than any LLM judge we tested".

[B] Simon Willison — Jev introduces a new shape of LLM—System One, aka Decision Models (2026-09-21)

https://simonwillison.net/2026/Sep/21/jev/

• The input "state" is "a string, array of strings, or set of name-value pairs", and the output is only "floating point numbers" with "associated confidence scores".
• Question types: Noul ("short for Bernoulli", "a floating point number between 0 and 1"), Choice ("a confidence score plus a probability distribution across all of the options") and Score ("numeric levels with descriptions").
• He sides with Maggie Appleton that "decision models" is "a better name for these".
• Questions "are evaluated in parallel, so sending many questions should take a similar time to sending just one".
• Only input is billed, at "$0.042 per million tokens", "cheaper even than OpenAI's GPT-5 Nano ($0.05/million)".
• In one experiment Jev scored every Bay Area city on "Good city?", and "it rated Cupertino top and East Palo Alto bottom".
• Good fits are "spam detection, suggesting labels, prioritization and ranking", plus search re-ranking that fetches candidates with BM25 and has Jev score them.
• Poor fits are "numbers, dates, or 'adversarial content'"; it is also a black box that gives no reasons, and bias is a concern in high-stakes uses such as ranking job applicants.
• He released llm-typesafe, a plugin that "adds support for Jev" to his LLM CLI tool.

[B] Latent Space (AINews) — Here are 6 Clones of Jev in 2 days (2026-09-19)

https://www.latent.space/p/ainews-here-are-6-clones-of-jev-in

• Six clones in two days: Laya (421M params, ModernBERT-large with two added transformer layers, PPO), DiffusionGemmaJev (diffusion-based, vLLM PR #57250), Bespoke Nimble (LoRA on Qwen3.5-9B, synthetic data, 66% to 90%), SemIf (formerly OpenJev; a three-class NLI head on Qwen3.5 4B / 35B), Jevlike (a lightweight option-attention model with 40K byte embeddings) and Kev-0.5B (LoRA plus a readout head on Qwen2.5-0.5B, runs on a MacBook Pro).
• The launch video drew "36M views" in two days, and "discriminative models broke out as a new systems primitive".
• Caveats: the data is "acknowledged to be 100% synthetic", "there is still no standard benchmark for this category", and demos "emphasized speed more than quality".
• Reactions split: "post-ChatGPT users treated it like a revelation, while pre-GPT ML people were more puzzled by the hype".

[B] Flavio Copes — A deep dive into Jev, TypeSafe's System One model (updated 2026-09-24)

https://flaviocopes.com/jev/

• Jev launched on 09-15, and by 09-22 TypeSafe had "paused new signups".
• Where an LLM "generates its answer one token at a time", Jev "samples all the answers in parallel" against the same state.
• Because "a response cannot contain a value outside your schema", "schema errors and wrong decisions become separate problems".
• The published figures (70 to 500 ms end to end, "most around 100", $0.042 per million input tokens, 40x to 200x faster) are TypeSafe's own, and its headline multiples are best treated "as a ceiling".
• The author has "console access" but has not put Jev "into production yet", and there is no independent accuracy check.
• Good fits: routing, labelling at scale, a safety gate that judges the risk of shell commands, search re-ranking, and turning free text into features for classical ML.
• Poor fits: arithmetic on numbers and dates, generation or summarisation, images and audio, and extraction where the model would have to invent the value.
• "Where deterministic code already decides correctly, code stays."
• Open questions: whether answers "stay accurate after the model changes", the "cost per solved task, not per token" once retries and human review are counted, how answers hold up against hostile inputs when the 0% type-error rate "covers the shape of an answer, not whether it's correct", and whether decision models become "a larger market than generative LLMs".

[B] DataCamp — Jev: TypeSafe's System One Model That Never Hallucinates (2026-09-16)

https://www.datacamp.com/blog/system-one-models-jev

• "Never hallucinates" means that no value outside the schema can come out, because "valid outputs are defined in the schema in advance".
• TypeSafe's benchmark on 4 business workflows (accuracy / cost per case / latency): Jev 67.8% / $0.0004 / 0.4 s, GPT-5.6 Terra 67.9% / $0.0304 / 10.1 s, GPT-5.6 Sol 74.1% / $0.0836 / 23.3 s, Claude Opus 5 73.1% / $0.1761 / 37.8 s.
• The $0.042/M input price is "roughly 1/48th of GPT-5.6 Terra's $2.00 per million input tokens".
• The endpoint is POST https://api.typesafe.ai/v1/systemone.
• The split puts Jev on "the high-volume routing layer" and escalates to an LLM when "a written explanation" is needed, and calibrated probabilities mean "you can route on a confidence threshold".

[B] laiso — #21 Jev (2026-09-17, in Japanese) / A comparison using System One Adapter (2026-09-25, in Japanese)

https://blog.lai.so/2026-09-17/ / https://blog.lai.so/system-one-adapter/

• The input is the state or context used for the decision, the question type, and the criteria or options; the output is the chosen answer, a rating, probabilities and a confidence number.
• Demos of Jev playing Super Mario and flying a drone simulator drew attention.
• On Vercel AI Gateway, Jev was used by more paying teams in its first 24 hours than any model before it.
• As of 09-25 new signups were closed, and the author, aiming to build a Jev-compatible model, set up an environment that compares the Jev API with local models on the same tasks (the results are in the paid section).

[C] YESOD ik-fib — Trying Jev, an AI that returns no prose (2026-09-25, in Japanese)

https://zenn.dev/yesodco/articles/yesod-jev-typesafe-inquiry-triage

• An experiment that routes internal inquiries to three paths: pointing to the documentation, a first check by CS, or escalation to development.
• Five questions: category (choice, 5 classes), answerable_from_docs (noul), needs_code_or_log (noul), cs_can_verify (noul) and urgency (score, 0 to 2).
• 8 inquiries cost about $0.0004 (about $0.00005 each); latency was not measured, but it felt faster than an LLM.
• Because the response format is fixed, there are no parse failures.
• One classification came back at 49% confidence, so low-confidence handling should be designed up front; two of the four designed branches never fired; and Japanese accuracy trails English (per the official documentation).
• The model is jev-latest, authentication uses TYPESAFE_API_KEY, and calls are plain HTTP POSTs with the standard library.

[C] Zenn, HN and Ollaya (feed) — Community reactions (2026-09-17 to 2026-09-25, partly in Japanese)

• Since mid-September the timeline has been filled with Jev (Zenn, 30+ users on Hatena Bookmark).
• A Zenn post argues against making Jev play games and for distilling an LLM instead: games whose input is not natural language can be handled by decision trees or small distilled models.
• Ollaya, an open-source project to "Run decision models locally", has appeared.
• HN (116 points): a discussion of "LLM Classification Is Feature Engineering".

Notes

• All performance and cost figures are TypeSafe's own claims, and there is no independent accuracy check yet (Flavio Copes, AINews).
• "Never hallucinates" is about format: the 0% type-error rate "covers the shape of an answer, not whether it's correct" (Flavio Copes), while DataCamp presents it as a strength.
• The name wavers between "System One model" and "decision model".
• The official documentation says accuracy in Japanese trails English (YESOD).
• New signups have been paused since 09-22, and when they will reopen is unverified.
• The claim that Jev was used more than any other model within 24 hours comes from laiso and does not appear in Vercel's changelog (unverified).

Ideas for practice and output

• In an LLM app of your own, pick one place where an LLM does the classifying (routing, guards, labelling) and rewrite it as Choice / Noul questions.
• Until signups reopen, build the same evaluation setup with a local decision model such as Ollaya or OpenJev, and put the judgements and probabilities of an LLM judge, the local decision model and Jev side by side.
• Route cases into three tiers by probability threshold (automatic / escalate to an LLM / human review) and measure the error rate at each threshold.
• Compare miscalibration on Japanese input with English input on the same material.

3. 📄 Claude Code started reading AGENTS.md, moving the standardisation and upkeep (prompt-audit) of agent instruction files a step further

• Selected / impact 2 / potential 2 / anchor 09-18 / track: coding AI.
• Feed items in this cluster:
  • claude-code-changelog 09-18 2.1.277
  • hn-claude-code 09-18 Claude Code now reads AGENTS.md if there is no Claude.md (740 points)
  • hn-claude-code 09-23 Claude Code reads AGENTS.md only when telemetry is on [fixed] (484 points)
  • publickey 09-23 Claude Codeが「AGENTS.md」に対応 (Claude Code now supports AGENTS.md)
  • publickey 09-14 「Agent Router」、Linux Foundation傘下で業界標準へ (Agent Router heads for industry-standard status under the Linux Foundation)
  • claude-code-changelog 09-25 2.1.283
  • classmethod-claude-code 09-26 Claude Code v2.1.282〜v2.1.283 の主要アップデート (key updates in Claude Code v2.1.282 to v2.1.283)
  • zenn-trending 09-25 /claude-api prompt-audit で棚卸ししたら (taking stock with /claude-api prompt-audit)

Collected information

[A] Anthropic — Claude Code changelog 2.1.277 (2026-09-18)

https://code.claude.com/docs/en/changelog#2-1-277

• "Added AGENTS.md support: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead; change it under 'Project instructions' in /config".
• The same release added CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 "for Claude apps gateways whose only egress is a forward proxy".
• At release the entry ended with "(not yet on Bedrock, Vertex or Foundry)"; the changelog update of 2026-09-25 removed it.

[A] Anthropic — Claude Code changelog 2.1.281 (2026-09-23)

https://code.claude.com/docs/en/changelog#2-1-281

• "Changed AGENTS.md support to also work on Amazon Bedrock, Google Vertex AI, Microsoft Foundry, LLM gateways, and sessions with telemetry disabled".
• The line was added to the changelog on 2026-09-25 (commit 163ae3a264 in anthropics/claude-code), after the Publickey and Szypowicz posts.

[A] Anthropic — Claude Code changelog 2.1.283 (2026-09-25)

https://code.claude.com/docs/en/changelog#2-1-283

• Added /doctor prompt-audit, which audits CLAUDE.md files for "prompting patterns written for older models".
• Added the availableModelsMatch managed setting (with "exact", an availableModels entry "allows only the model version it names") and the deniedModels managed setting.

[B] Publickey — Claude Code now supports AGENTS.md (2026-09-24, in Japanese)

https://www.publickey1.jp/blog/26/claude_codeagentsmdclaudemd.html

• In a project with no CLAUDE.md, Claude Code reads AGENTS.md automatically, and this can be changed under "Project instructions" in /config.
• It is not yet supported on Amazon Bedrock, Google Vertex or Microsoft Foundry.
• OpenAI Codex, Google Gemini CLI, Devin, Cursor, GitHub Copilot and others already support AGENTS.md, and this is the first time Claude Code does.

[B] Przemysław Szypowicz — Claude Code reads AGENTS.md only when telemetry is on (2026-09-23)

https://blog.szypowi.cz/p/claude-code-reads-agents.md-only-when-telemetry-is-on/

• With DISABLE_TELEMETRY=1 or CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 set, "AGENTS.md never loaded", even on 2.1.277 and later.
• The author checked this with "an empty directory that holds only an AGENTS.md with a canary word in it" (PERIWINKLE).
• The cause: the agents-md loader "asks a remote feature flag called tengu_agents_md_mod, with false as the fallback", and with telemetry off the flag cannot be fetched.
• The post points to GitHub issue #95690.
• It argues that "a privacy setting should never quietly switch off unrelated local behavior", and that the gate hits "teams on Bedrock, Vertex or a gateway" especially hard.
• The HN submission title carries [fixed]; the changelog lists the fix under 2.1.281.

[C] Hacker News — Claude Code now reads AGENTS.md if there is no Claude.md (2026-09-18, 740 points)

https://news.ycombinator.com/item?id=49760187

• Against the rule that CLAUDE.md takes precedence, many commenters want full parity with the standard that the other vendors follow.
• Workarounds such as ln -s AGENTS.md CLAUDE.md and git hooks have long been in use, hence the question "why did it take so long?"
• One view is that it took "the Shopify CEO tweet and Astra causing a large move towards Codex" for Anthropic to act.
• Others point out that Claude Code "still won't detect skills on .agents/skills", so its support for the standard is incomplete.
• Criticism of lock-in was widespread.

[B] Publickey — Agent Router heads for industry-standard status under the Linux Foundation (2026-09-15, in Japanese)

https://www.publickey1.jp/blog/26/openaianthropicapiagent_routerlinux_foundation.html

• The Agentic AI Foundation (AAIF) is an organisation under the Linux Foundation that drives the standardisation of MCP, AGENTS.md, the Agent2Agent protocol and related technologies.
• Agent Router (formerly Envoy AI Gateway) is an implementation layer that absorbs the differences between vendor APIs and unifies multiple MCP servers behind a single endpoint; version 1.1 is now production-ready.
• Supported vendors include OpenAI, Anthropic, Amazon Bedrock, Google Gemini, Vertex AI, Azure OpenAI, xAI, Groq, Mistral, Cohere, DeepSeek and Together AI.

[B] Classmethod — Key updates in Claude Code v2.1.282 to v2.1.283 (2026-09-26, in Japanese)

https://dev.classmethod.jp/articles/20260924-cc-updates-v2-1-283/

• On a sample project, /doctor prompt-audit (/checkup prompt-audit also works) flagged a nonexistent path, an undefined command, a contradiction between instruction files (the commit-message language), leftover workarounds for older models (a word limit, an instruction to plan) and excessive emphasis.
• Auto mode becomes the default: with a third-party provider or with telemetry off, a session with no permission mode set starts in auto mode, and permissions.defaultMode still takes precedence.
• Administrators get deniedModels and availableModelsMatch: "exact".
• There are 9 breaking changes, including OpenTelemetry variables in project settings now being ignored, sandbox.excludedCommands being ignored under managed settings, and the anthropic-skills namespace being reserved.
• Managed-settings values that used to be ignored because of an invalid entry start to take effect after the update, so the article recommends checking the startup warnings.

[C] Zenn nanora — Taking stock with prompt-audit (2026-09-25, in Japanese, feed)

• Run on the author's own agent, it turned up many mismatches between what is written and what actually exists, rather than problems of wording.
• opus had quietly become Opus 5.5, and the effort setting was not taking effect, so it stayed on the heavy xhigh.
• Before deleting an instruction, compare behaviour with and without it (deleting one made things clearly worse).

Notes

• The telemetry-off fix and support on Bedrock / Vertex / Foundry both sit under 2.1.281 (09-23), but only since the changelog update of 09-25.
• Publickey's "not yet on Bedrock / Vertex / Foundry" reflects the 2.1.277 entry at release and is out of date since 2.1.281.
• .agents/skills is still unsupported, according to an HN comment.
• The AGENTS.md specification itself (agents.md) was not read in this run.

Ideas for practice and output

• In your own global configuration (a CLAUDE.md that imports AGENTS.md), check how Claude Code behaves when CLAUDE.md is removed and only AGENTS.md is left, with telemetry both on and off.
• Run /doctor prompt-audit on the global CLAUDE.md / AGENTS.md and on one project, sort the findings, and decide what to keep and what to remove.
• Try availableModelsMatch in managed settings and confirm a setup in which new models do not automatically become the default.

4. 🛡️ Copilot, Kiro, Claude Code and Cursor shipped controls for unattended agent runs in the same week

• Selected / impact 2 / potential 2 / anchor 09-19 to 09-25 / track: coding AI.
• Feed items in this cluster:
  • github-changelog-copilot 09-23 Local sandboxing in the GitHub Copilot app
  • github-changelog-copilot 09-23 OpenTelemetry in the GitHub Copilot app
  • github-changelog-copilot 09-23 New features and improvements in Copilot for JetBrains (AI-assisted tool approvals)
  • kiro-changelog 09-23 CLI 2.24: Session-Wide Tool Approval, Session Dashboard Controls, and Environment Isolation
  • claude-code-changelog 09-19 2.1.278
  • claude-code-changelog 09-23 2.1.281
  • claude-code-changelog 09-24 2.1.282
  • claude-code-changelog 09-25 2.1.283
  • cursor-changelog 09-23 Rollouts and Security Review
  • classmethod-claude-code 09-24 v2.1.281 の主要アップデート (key updates in Claude Code v2.1.281)
  • classmethod-claude-code 09-25 Claude初心者がAutoModeでも安心できるセキュリティ設定をしてみた話 (a Claude beginner tries security settings for using AutoMode with peace of mind)
  • classmethod-claude-code 09-24 settings.json を Claude Code 自身に書き換えさせない方法 (how to stop Claude Code from rewriting its own settings.json)
  • classmethod-claude-code 09-25 Claude の自動実行が権限不足のエラーが出るときの原因と対処法 (why Claude's automatic runs fail with insufficient-permission errors, and how to fix them)

Collected information

[A] GitHub — Local sandboxing in the GitHub Copilot app (2026-09-23)

https://github.blog/changelog/2026-09-23-local-sandboxing-in-the-github-copilot-app

• It restricts three areas: the filesystem (additional read/write, additional read-only and denied folder lists), the network (outbound internet and local network) and credentials (Git credentials for authenticated HTTPS git operations, and GitHub CLI credentials for GitHub CLI authentication).
• It is configured per project under "Sandbox" in the app settings, and the effective policy "can be more restrictive when enterprise-managed settings apply".
• /sandbox on enables sandboxing for an active local session.
• "If your operating system cannot enforce the requested policy, the sandboxed shell fails with an error rather than running without a sandbox."
• It is off by default, applies only to new sessions (or when an existing session restarts), does not apply to cloud sandbox sessions or sessions running on a remote host, and is in public preview.

[A] GitHub — OpenTelemetry in the GitHub Copilot app (2026-09-22)

https://github.blog/changelog/2026-09-22-opentelemetry-in-the-github-copilot-app

• Teams can follow the flow of a session, including requests to AI models and the tools an agent uses, and review step-by-step traces of agent execution.
• The telemetry property in managed-settings.json specifies the endpoint that will receive the data.
• Prompt and response content is excluded by default.
• The demo pairs it with Splunk Observability Cloud.

[A] Kiro — CLI 2.24.0: Session-Wide Tool Approval, Session Dashboard Controls, and Environment Isolation (2026-09-23)

https://kiro.dev/changelog/cli/2-24

• /tools trust-all automatically approves tool calls for the rest of the current V3 session (after a safety warning and a confirmation; enable it "only in an environment you trust").
• Project .env files no longer load automatically into chat sessions, MCP servers or tools, and Kiro starts with just the environment inherited from the terminal.
• Where a workflow needs those variables, "export them in your shell before starting Kiro", and MCP servers reference the exported values in their configuration.
• In /sessions, the Filter, Sort and Group controls show their options inline, and the filters can be combined.

[A] Anthropic — Claude Code changelog 2.1.278 / 2.1.281 / 2.1.282 / 2.1.283 (2026-09-19 to 2026-09-25)

https://code.claude.com/docs/en/changelog

• 2.1.278: auto mode for Claude API and Enterprise users, and on Bedrock, Vertex, Foundry and gateways, now defaults to the server-side classifier, which does not charge for classifier overhead (CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out).
• 2.1.281: fixed a recursive rm whose target is only command-substitution output running unprompted in auto mode.
• 2.1.281: fixed a turn that could retry indefinitely, ignoring --max-turns, and the prompt cache being lost when an MCP server disconnects.
• 2.1.281: the gateway configuration gains assume_role on Bedrock upstreams (the gateway assumes an IAM role through STS, optionally one session per developer) and guardrail: {id, version} (applies an Amazon Bedrock guardrail to every request).
• 2.1.281: "attribution": false in settings.json hides all commit and PR attribution.
• 2.1.281: the dangerous rm prompt in --dangerously-skip-permissions and auto mode now waits 2 minutes for an answer, then denies the command with a rewrite hint (CLAUDE_CODE_DISABLE_DANGEROUS_RM_TIMEOUT=1 turns this off).
• 2.1.282: auto mode also defaults to the server-side classifier when telemetry is off.
• 2.1.282: a startup notice and entries in /status and claude doctor list telemetry variables in a project's settings files that were ignored or that turned telemetry off.
• 2.1.283: x-claude-code-prompt-id joins the gateway hint headers (CLAUDE_CODE_GATEWAY_HINT_HEADERS=1).
• 2.1.283: availableModelsMatch: "exact" and deniedModels.

[A] Cursor — Rollouts and Security Review (2026-09-23)

https://cursor.com/changelog/rollouts-and-security-reviewer

• Rollouts attaches a monitor to every pull request and "watches the change as it deploys, reporting change health per environment: verified healthy, regression detected, or inconclusive".
• It reads the diff and the systems it touches, writes a monitoring plan as a PR comment, and checks the change's intended effect alongside error and latency signals.
• When it suspects a regression it notifies the author and, depending on configuration, can open a revert PR or hand the finding to a cloud agent for a fix, but it does not merge or roll back on its own today.
• It connects to Origin or GitHub for source control, to the continuous delivery system for deploy events and to Datadog and other telemetry providers for signals, and feature flag integration is coming soon.
• Security Review reads every pull request in the context of the codebase and posts "one review comment reporting exploitable bugs".
• It looks for injection across SQL, command and template surfaces, authentication and authorisation bypasses, secrets committed to source, SSRF, unsafe deserialisation and dependency changes that introduce known vulnerabilities, and each finding carries the attack path and a proposed fix.
• Draft PRs are skipped, and a finding dismissed with a reason is not raised again on that PR.
• Both bots are for Teams and Enterprise plans and are enabled from the automations tab, with usage credits for the next 10 days to try Rollouts (roughly 50 changes on Teams, 500 on Enterprise).
• "Style and quality stay with Bugbot."

[B] Classmethod — Key updates in Claude Code v2.1.281 (2026-09-24, in Japanese)

https://dev.classmethod.jp/articles/20260924-cc-updates-v2-1-281/

• A recursive rm whose target is only command-substitution output, such as rm -rf "$(pwd)", no longer runs without confirmation in auto mode and with --dangerously-skip-permissions.
• Where the classifier review runs server-side, read-only shell commands and sandboxed commands now also wait for that review.
• The author confirmed that commits carry Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> by default and that "attribution": false drops the trailer.
• To deal with unexpected denials in unattended runs, the author wants to be able to trace them in logs.

[B] Classmethod — Security settings for using AutoMode with peace of mind (2026-09-25, in Japanese)

https://dev.classmethod.jp/articles/claude-security-configuration-tips/

• hooks: a script runs just before each command, blocking git reset --hard, git push --force and the like with deny and letting everything else through.
• permissions: commands are sorted into allow / ask / deny, with git commit and git push set to ask to force a visual check.
• Bash sandbox: access to .env and similar files is restricted at the OS level, blocking access to secrets across every project under the development directory.
• Limits: dynamic shell variable expansion such as CMD="reset --hard"; git $CMD cannot be caught by static analysis, which the author calls a limit in principle, and covering every deletion command other than git and rm is not realistic.
• Conclusion: decide in advance what to leave to the AI, where a human steps in and what the system blocks.

[B] Thoughtworks — As AI Accelerates Software Complexity, Thoughtworks Technology Radar Urges a Return to Engineering Fundamentals to Combat Cognitive Debt (2026-04-15, background, outside the window)

https://www.thoughtworks.com/about-us/news/2026/combat-ai-cognitive-debt-radar-v34

• Two of the themes of Technology Radar volume 34 are "Securing permission-hungry agents" and "Putting coding agents on a leash".

[C] Classmethod, Kiro and Copilot (feed) — Related items (2026-09-23 to 2026-09-25, partly in Japanese)

• In the same week Classmethod also published how to protect settings.json with protected paths and deny rules, and how to fix permission errors in automatic runs.
• Copilot for JetBrains 1.18.0 brings "AI-assisted tool approvals" and shared skills and instructions for the organisation (feed).
• Kiro CLI 2.23 (09-21) lets you configure thinking and effort from /model and start V3 cloud sessions before connecting a repository (feed).

Notes

• Copilot's sandbox is off by default and in preview, and it does not cover cloud sandbox sessions or remote hosts.
• In 2.1.278 and 2.1.282 Claude Code widened where auto mode is set to "default to the server-side classifier", so which environments are billed depends on the changelog's wording.
• Cursor's two bots are limited to Teams and Enterprise, so they cannot be tried on an individual plan.

Ideas for practice and output

• For each task that runs unattended, write down its permitted scope on three axes (files, network and credentials), using Copilot's categories as the checklist.
• For unattended Claude Code runs, pin the model in managed settings and, taking the auto mode classifier as a given, make denials traceable in logs.
• Switch to a setup that no longer passes .env implicitly to MCP servers and tools.
• List the destructive commands that hooks should deny, and back the list with the sandbox, given the limits on catching dynamic expansion.

5. 🏗️ Azure, Google and AWS shipped platform controls for always-on agents (triggers, egress control, anomaly detection, multi-account access, a new runtime)

• Selected / impact 2 / potential 2 / anchor 09-15 to 09-24 / track: GenAI engineering.
• Feed items in this cluster:
  • microsoft-foundry-blog 09-24 Routines in Microsoft Foundry Are Now Generally Available
  • microsoft-foundry-blog 09-24 Control where your hosted agent connects with network egress
  • google-developers-blog (no date) Agent Anomaly Detection, now in Private Preview
  • google-developers-blog (no date) Build zero-trust AI agents that judge intent, not just syntax
  • aws-ml-blog-bedrock 09-24 Build a multi-account AI agent with AgentCore Gateway and MCP
  • aws-whats-new 09-18 The new AgentCore Runtime is now available
  • aws-jp-weekly-genai 09-24 週刊生成AI with AWS – 2026/9/14 週 (Weekly Generative AI with AWS, week of 2026-09-14)
  • langchain-blog 09-24 New in LangSmith Engine: red teaming and automated testing (peripheral)

Collected information

[A] Microsoft — From Chatbots to Automated Agents: Routines in Microsoft Foundry Are Now Generally Available (2026-09-24)

https://devblogs.microsoft.com/foundry/from-chatbots-to-automated-assistants-routines-in-microsoft-foundry-are-now-generally-available/

• "The first generation of agents waited for a person to send a message", whereas organisations now want agents that "monitor work, respond to business events, and continue tasks over time".
• Routines have three trigger types, Timer (once at a future date and time), Recurring (repeatedly on a schedule, set with a CRON expression) and Event-based (GitHub issue events, new messages in Microsoft Teams channels); a separate reminder tool lets a hosted agent "schedule itself to run again".
• Routines run agents in Foundry Agent Service and are created with the Python SDK (azure-ai-projects), REST, the JavaScript and .NET SDKs, the Azure Developer CLI or the portal.
• Run history and the audit trail live in the Foundry project, and a run acts either with "the creator's delegated access" or as the agent, using "its own configured permissions in Microsoft Entra ID".
• Example uses: GitHub issue triage, daily summaries and weekly reports, compliance checks, answering support requests in Teams channels, and checking back on a long-running process until it finishes.
• No pricing or limits are stated.

[A] Microsoft — Control where your hosted agent connects with network egress in Foundry Agent Service (2026-09-24)

https://devblogs.microsoft.com/foundry/egress-controls-hosted-agent/

• An explicit destination policy decides where a hosted agent may connect, in one of two modes: Audit (nothing is blocked; calls that would be denied are only recorded, as "would-deny evidence") and Enforced (a destination outside the allow list gets a "Proxy 403").
• The policy is an ARM resource written in JSON, with mode: Audit|Enforced, defaultAction: Deny and ruleType: Fqdn rules.
• The policy is created with the Azure CLI (az rest) and attached with RaiConfig in the Python SDK, or created and assigned in the portal.
• Decisions are observed in "the invocation's network egress decision spans or its project's Application Insights records".
• The motivation is to "keep destination rules outside your agent code" and to guard against the case where "a document includes an unfamiliar upload link, or a new helper library follows a URL you did not expect".
• It is a preview: "Network egress controls are not GA, have no preview SLA, and are not intended for production use."

[A] Google — Agent Anomaly Detection, now in Private Preview on the Gemini Enterprise Agent Platform (2026-09-16)

https://developers.googleblog.com/agent-anomaly-detection-now-in-private-preview-on-the-gemini-enterprise-agent-platform/

• A "reasoning-based oversight and audit layer for autonomous agents" whose analysis "runs asynchronously and out of band from the live request path".
• It reads "the logs and OpenTelemetry traces your agents already emit": a lightweight first pass surfaces statistical anomalies, then "an LLM-based reasoning layer deeply examines the flagged sessions".
• Detected risks map to the OWASP Agentic Top 10 (tool misuse, identity and privilege abuse, cascading failures, rogue agents), plus resource exhaustion and token usage escalation (for example, "systematic scraping" through large-batch tool calls).
• Provisioning is one click with ADK 1.2 or later, findings appear in Security Command Center with "a severity, a plain-language explanation of what triggered it, and recommended next steps", and an API lets an ADK callback or plugin respond to them automatically.
• In Private Preview.

[A] Google — Build zero-trust AI agents that judge intent, not just syntax (2026-09-15)

https://developers.googleblog.com/build-zero-trust-ai-agents-that-judge-intent-not-just-syntax/

• Explains the move from static, build-time security controls to dynamic runtime governance through three managed controls.
• Model Armor screens the payload "at the ingress perimeter" for "prompt injection, jailbreaks, and malicious URLs", and at egress runs Sensitive Data Protection to redact outgoing responses.
• Semantic Governance Policies are "an LLM-based natural-language policy engine that evaluates each proposed tool call against user intent and your business rules before the call runs".
• Agent Anomaly Detection "monitors session telemetry across the fleet, using statistical models and LLM analysis", catching multi-step abuse such as "a refund drained across many turns".
• Rules are "written as plain-text constraints" (for example, "Refunds for opened digital goods, software licenses, or clearance items over 30 USD must be denied and routed to a human manager").
• Remediation closes the loop "without modifying or redeploying agent code": from an anomaly finding, an administrator reviews the flagged trace and "authors a new natural-language constraint that covers the multi-turn pattern", a step that "can also be handled programmatically via APIs".

[A] AWS — Build a multi-account AI agent with AgentCore Gateway and MCP (2026-09-24)

https://aws.amazon.com/blogs/machine-learning/build-a-multi-account-ai-agent-with-agentcore-gateway-and-mcp/

• Hub-and-spoke: a central platform account runs the agent on AgentCore Runtime with inference on Bedrock, line-of-business (LOB) accounts expose their data as MCP servers (FastMCP) on AgentCore Runtime, and AgentCore Gateway in the platform account "aggregates them behind a single endpoint".
• Authentication: the agent forwards the user's JWT to the Gateway, the Gateway retrieves OAuth 2.0 M2M credentials from AgentCore Identity, each LOB MCP server authenticates the inbound token against Okta's OIDC endpoint, and the user's JWT travels in the Authorization header "for identity propagation".
• Services used: AgentCore (Runtime, Gateway, Identity, Evaluations, Optimization), Strands Agents, Bedrock and Knowledge Bases, DynamoDB, S3 with OpenSearch Serverless, ECS with Fargate, CloudFront, Okta and CloudWatch.
• "The LOB's data stays in its own account"; Policy in AgentCore (Cedar) on the Gateway enforces least privilege with default-deny, PII redaction and content filtering are applied centrally, and CloudTrail audits across accounts.
• Constraints: the reference implementation uses AgentCore Runtime's public network mode, so production needs VPC connectivity through ENIs; it does not use OAuth on-behalf-of (OBO) token exchange because "the Okta developer account used for the sample doesn't support the OBO flow"; and it assumes a multi-account setup under AWS Organizations.
• Cost: inference and Gateway invocations "accrue in the platform account" and are apportioned with cost allocation tags, while each LOB's MCP server and data-plane costs stay in its own account.

[A] AWS — The new AgentCore Runtime is now available in Amazon Bedrock AgentCore (2026-09-18)

https://aws.amazon.com/about-aws/whats-new/2026/09/new-agentcore-runtime-generally-available

• The next-generation AgentCore Runtime (serverless microVMs) has "elastic memory management" that "reclaims unused memory throughout the session so you pay for actual usage rather than the peak".
• It delivers "consistent cold start times regardless of container image size or concurrency", with a P75 of 1.9 to 2.0 seconds (5.4 to 30 seconds with V1).
• Regions: us-east-1 and us-east-2 (N. Virginia, Ohio), us-west-2 (Oregon), eu-west-1 (Ireland) and ap-northeast-1 (Tokyo).
• To migrate, set platformVersion to V2 when creating or updating a runtime.
• It was tested with 200 MB to 2 GB images, but no upper limit is stated.

[A] AWS Japan — Weekly Generative AI with AWS, week of 2026-09-14 (2026-09-24, in Japanese)

https://aws.amazon.com/jp/blogs/news/weekly-genai-20260914/

• Tokyo is among the Regions where the next-generation AgentCore Runtime is available.
• Moonshot AI's Kimi K3 (2.8 trillion parameters, 1M context) is generally available on Bedrock.
• In Kiro, the GPT-5.6 family goes from 272K to 1M tokens, and Claude Fable 5.1 is in preview for Kiro Enterprise (1M context, 6x credit multiplier).
• In AWS's Deception Benchmark (14,822 samples, 16 languages), precision with the standard prompt was in the mid-50% range, and no configuration among 12 models from 5 vendors kept both false positives and false negatives under 10%.

[C] LangChain (feed) — LangSmith Engine v2: Red teaming and automated testing (2026-09-24)

• LangSmith Engine "now includes Red Teaming to proactively detect agent issues and automated agent testing" (peripheral information; body not fetched).

Notes

• The two Google posts carry no date in the feed, so their anchors were judged from the dates on the pages (09-15, 09-16).
• No independent analysis covering the three vendors together was found inside the window (unverified).
• The AWS reference implementation spells out constraints that keep it from production use as is (public network mode, no OBO support).
• Foundry Routines pricing and limits are not stated.

Ideas for practice and output

• List your agent's outbound destinations, observe them first in an Audit-like, log-only mode and only then switch to Enforced (copy Foundry's egress policy JSON by hand and write an allow list for your own agent in the same format).
• Run one AgentCore Runtime V2 in the Tokyo Region and measure its cold starts and how its usage shows up in billing.
• Write one Cedar default-deny policy for AgentCore Gateway and check that only the permitted tools get through.
• Emit OTel traces per tool call, at a granularity that anomaly detection can take as input.

6. 🔌 GCP and AWS shipped features that turn existing REST APIs and SaaS knowledge into agent tools through MCP and Agent Skills

• Dropped / impact 1 / potential 2 / anchor 09-23 to 09-25 / track: GenAI engineering.
• Feed items in this cluster:
  • google-developers-blog (09-24) Turn your REST APIs into MCP tools with Google Cloud API Gateway
  • aws-whats-new 09-25 AWS End User Messaging and Amazon SES now offer AI agent skills for the AWS MCP Server
  • aws-whats-new 09-23 Amazon Bedrock Managed Knowledge Base now supports Salesforce and Zendesk
  • aws-ml-blog-bedrock 09-23 HEMA's journey with MCP and Amazon Bedrock
  • hn-mcp 09-20 MCP was always a bad idea? (334 pt, a counterpoint)
• Reason for dropping it: each feature is small and falls short of the top five (re-evaluate next time on whether MCP in API Gateway leaves preview or AWS adds more Skills).

Collected information

[A] Google — Turn your REST APIs into MCP tools with Google Cloud API Gateway (2026-09-24)

https://developers.googleblog.com/turn-your-rest-apis-into-mcp-tools-with-google-cloud-api-gateway/

• Annotating an OpenAPI 3.0.x / 3.1.x spec with x-google-api-management.mcp: true (document level) and x-google-mcp-tool (operation level) makes the gateway accept MCP JSON-RPC requests on a single /mcp endpoint and transcode each tools/call into the corresponding REST request.
• "Your existing spec, gateway, authentication, quotas, and logging do the work."
• Authentication is by JWT or API key on tools/call, while for tools/list in production the post says to "require a JWT".
• A Python example with ADK's McpToolset and StreamableHTTPConnectionParams.
• In Public Preview, operations returning HTTP 204 are not exposed, a gateway serves up to 1,000 tools, and MCP and model routing cannot be enabled in the same API config, while "MCP resources and prompts, response streaming, and Model Armor payload inspection are on the roadmap".

[A] AWS — AWS End User Messaging and Amazon SES now offer AI agent skills for the AWS MCP Server (2026-09-25)

https://aws.amazon.com/about-aws/whats-new/2026/09/aws-messaging-ses-ai-skills-mcp-server/

• In Claude Code, Codex and Cursor the aws-core plugin bundles the AWS MCP Server, while in Kiro and other agents "you add the server to your MCP configuration file".
• The SES skill lets you ask the agent to "verify a sending identity and send your first production email", and the End User Messaging skill to "build a branded RCS agent and send yourself a rich message", with SMS / WhatsApp covered as well.

[A] AWS — Amazon Bedrock Managed Knowledge Base now supports Salesforce and Zendesk as native data source connectors (2026-09-23)

https://aws.amazon.com/about-aws/whats-new/2026/09/amazon-bedrock-managed-knowledge-base-salesforce-zendesk-native-data-source-connectors/

• Connectors that sync Salesforce knowledge articles and Zendesk help center articles and community posts into "a fully managed retrieval-augmented generation (RAG) service".
• The connectors "handle data crawling, metadata extraction, and incremental sync automatically".

[A] AWS — From portal-hopping to instant answers: HEMA's journey with MCP and Amazon Bedrock (2026-09-23)

https://aws.amazon.com/blogs/machine-learning/from-portal-hopping-to-instant-answers-hemas-journey-with-mcp-and-amazon-bedrock/

• HAL is a Strands agent hosted on AgentCore Runtime that draws on "several knowledge bases built on Amazon Bedrock Knowledge Bases" (IT documentation, API specifications, Kafka schemas, the service catalog).
• Two AgentCore Gateways, one for internal use (IAM SigV4) and one for external clients (Entra ID OAuth2 with custom JWT validation), make the tools usable from MCP clients such as the HAL web chat, Kiro and Claude.
• There are "no AWS credentials on the client", Active Directory groups control permissions on the read-only tools, and an MCP auth proxy runs on API Gateway + Lambda.
• No quantitative results are reported, and the lesson drawn is that "the read architecture earns the write step".

7. 🗣️ A debate over "development without hand-written code" spread from DHH's Rails World keynote

• Dropped / impact 2 / potential 1 / anchor 09-24 / track: coding AI.
• Feed items in this cluster:
  • pragmatic-engineer 09-24 The Pulse: RoR creator sparks new "death of coding by hand" debate
  • hatena-ai-agent 09-24 Rails World 2026 DHHキーノートの日本語要約 (a Japanese summary of DHH's Rails World 2026 keynote)
  • hatena-ai-agent 09-26 DHHはRailsを捨てたのか？ (asking whether DHH has abandoned Rails)
  • hatena-claude-code 09-24 ソフトウェアエンジニアとしてのぼくは、もう死んでいた (the software engineer in me was already dead)
  • martin-fowler 09-17 I don't like LLMs
  • pragmatic-engineer 09-15 Inside OpenAI's agentic software factory
  • hatena-llm 09-25 PS5非公式ハックの主導者が離脱、AI時代のOSS開発に苦言 (the leader of an unofficial PS5 hack steps away, criticising open-source development in the AI era)
• Reason for dropping it: the debate is broad, but as a change in practice it is one company's policy, with little technical substance.

Collected information

[C] komagata — A Japanese summary of DHH's Rails World 2026 keynote (2026-09-24, in Japanese)

https://komagata.github.io/public-notes/rails-world-2026/

• 37signals has stopped writing code by hand in its normal development work.
• DHH names Opus 4.5 as the turning point, after which he moved from detailed implementation instructions to stating the problem and the goal.
• An example is a mail server written in Rust, with CPU usage down 99% and memory down 95%.
• He advocates an asynchronous way of working that assigns tasks to agents instead of waiting for output in a chat window.

8. 🧰 Practical write-ups on harness design appeared one after another in Japan

• Dropped / impact 1 / potential 2 / anchor 09-16 / track: coding AI.
• Feed items in this cluster:
  • hatena-claude-code 09-16 Claude Codeで開発期間を2.5か月から1か月に縮めた「ハーネス」の設計手法 (how to design the "harness" that cut a development period from 2.5 months to 1 month with Claude Code)
  • hatena-ai-agent 09-24 ハーネス設計入門 〜 基礎知識の整理から実務へのステップアップ 〜 (an introduction to harness design, from organising the basics to stepping up to real work)
  • zenn-trending 09-25 人が押さえるべき5つの判断ポイントと半自動ループの作り方 (five decision points that humans should own, and how to build a semi-automated loop)
  • google-developers-blog (09-09, outside the window) The Anatomy of Harness Engineering
• Reason for dropping it: the in-window anchors are a case-study post and a conference talk deck, not a primary change to any specification.

Collected information

[A] Google — The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents (2026-09-09, background, outside the window)

https://developers.googleblog.com/the-anatomy-of-harness-engineering-how-to-evaluate-iterate-and-guard-ai-coding-agents/

• Evaluate: "behavioral evaluations" check intermediate steps rather than the final output and "function like integration tests for improving agent harness operation".
• Iterate: automate, with guardrails, "a loop where an LLM tweaks its own system prompt, iterating until a failing test finally passes".
• Guard: "automate batch evaluations to monitor stability" and let a CI/CD-style guardrail make sure changes "don't break any existing features", rather than "blocking PRs on single eval runs that can be noisy".
• A code example written for the Antigravity SDK, and mentions of Terminal-Bench and DeepSWE.

9. 🧪 A big release week for LangSmith

• Dropped / impact 1 / potential 1 / anchor 09-24 / track: GenAI engineering.
• Feed items in this cluster (langchain-blog 09-24 to 09-25): LangSmith Fine-Tuning and SmithTune, Managed Deep Agents 0.8 (user-owned credentials, per-user memory, an HTTP channel, file transfer in Slack, a web search tool), Trajectories (a conversation view of sessions), Engine v2 (red teaming and automated testing) and Custom Apps.
• Reason for dropping it: a platform update from one company, with limited impact beyond its users.

10. ⚙️ Operational changes to the Claude Platform API

• Dropped / impact 1 / potential 1 / anchor 09-14 to 09-24 / track: GenAI engineering.
• Feed items in this cluster (claude-platform-release-notes): 09-14 on-demand compaction in the Messages API (beta, compact-2026-09-04, signed compaction blocks), 09-18 the response format of cache diagnostics, 09-23 cache diagnostics GA (opt in with the diagnostics object), 09-24 billing resumed for refusals before any output (bio / frontier_llm / reasoning_extraction).
• Reason for dropping it: it only supplements theme 1.

11. 🧭 A run of Kiro updates

• Dropped / impact 1 / potential 1 / anchor 09-14 to 09-24 / track: coding AI.
• Feed items in this cluster (kiro-changelog): 09-14 IDE 1.1 (Agent Artifacts, native ARM64, Code OSS 1.131), 09-14 1M context for GPT-5.6 Sol / Terra / Luna, 09-16 CLI 2.22 (fullscreen chat), 09-16 an Enterprise preview of Claude Fable 5.1, 09-21 CLI 2.23 (reasoning controls, cloud sessions), 09-22 reasoning effort on the web, 09-24 IDE 1.1.70, 09-24 Crew 0.7.
• Reason for dropping it: apart from CLI 2.24, which is covered in theme 4, these are improvements to a single product.

12. 🏢 A batch of enterprise updates to GitHub Copilot

• Dropped / impact 1 / potential 1 / anchor 09-23 to 09-25 / track: coding AI.
• Feed items in this cluster (github-changelog-copilot): personal settings and enterprise defaults for code review, features enabled by default on Business / Enterprise (with a 28-day grace period), Slack / Teams integration, Copilot Memory in agentic autofix, a PR review stage in the usage metrics API, and a validator for managed settings.
• Reason for dropping it: apart from the sandbox and OTel items covered in theme 4, these are operational features.

13. 🎙️ Google model and platform updates

• Dropped / impact 1 / potential 1 / anchor 09-15 to 09-24 / track: GenAI engineering.
• Feed items in this cluster: deepmind-blog 09-15 Gemini 3.8 Live and 3.8 Live Extended Thinking, 09-23 Gemini 3.8 TTS, 09-23 server-side memory in Private AI Compute, 09-24 Gemini 3.8 Live with Live Avatar; gcp-agent-platform-release-notes 09-15 console support for RL fine-tuning (Preview), 09-16 / 09-21 / 09-24 CodeMender 0.8 to 0.10, 09-17 stateful video generation in Gemini Omni Flash, 09-18 Grok 4.6 GA and Agent Platform SDK for Python 2.0.1, 09-24 Gemini 3.8 Live GA.
• Reason for dropping it: centred on voice and video models, away from the core of this run's tracks.

14. 🧱 Open-weight models and inference efficiency

• Dropped / impact 1 / potential 1 / anchor 09-14 to 09-23 / track: GenAI engineering.
• Feed items in this cluster: aws-whats-new 09-18 Kimi K3 GA on Bedrock, 09-14 Qwen3.6-35B-A3B-NVFP4 and Wan2.1 on JumpStart, 09-16 serverless fine-tuning for Nemotron 3.5 Lightning; ainews 09-22 Xiaomi MiMo-V2.6-Pro (1T-A42B, $3M training cost); hatena-llm 09-23 notes on MoE fundamentals; hn-llm 09-23 Mercury 2.5 at 770 tok/s; aws-ml-blog-bedrock 09-23 a post on using OpenCode with open-weight models on Bedrock as a coding agent.
• Reason for dropping it: little connection to day-to-day practice in Japan.

15. ☁️ Claude Code cloud sessions made official

• Dropped / impact 1 / potential 1 / anchor 09-24 / track: coding AI.
• Feed items in this cluster: zenn-trending 09-24 「Claude Code クラウドセッション、使ってみて！」 (an invitation to try Claude Code cloud sessions, renamed from "Claude Code on the web" at the official release on 09-24), classmethod-claude-code 09-26 「Claude Code on the web の設定項目を整理する」 (sorting out the settings of Claude Code on the web), claude-code-changelog 09-14 2.1.271 (fast mode for Remote sessions).
• Reason for dropping it: mostly a rename and a tidy-up of settings, and no matching entry in the primary changelog has been confirmed.

16. 🔁 Codex model and CLI updates

• Dropped / impact 1 / potential 1 / anchor 09-14 to 09-26 / track: coding AI.
• Feed items in this cluster (openai-chatgpt-codex-changelog): 09-14 deprecation of GPT-5.3-Codex-Spark, 09-14 GPT-5.5 to retire from ChatGPT / ChatGPT Work / Codex on 10-14, 09-17 to 09-26 Codex CLI 0.155.0 to 0.157.1.
• Reason for dropping it: mostly notices of upcoming deprecations, relevant to a limited audience.

17. 📏 Practical material on evals and benchmarks

• Dropped / impact 1 / potential 2 / anchor 09-16 to 09-18 / track: GenAI engineering.
• Feed items in this cluster: hamel-husain 09-18 "AI Evals: Everything You Need to Know" (an FAQ drawn from teaching more than 5,000 people), generative-agents-blog 09-16 「AIエージェントキャッチアップ #91 - SkillsBench」 (AI agent catch-up #91 on SkillsBench, a benchmark for agent skills), aws-jp-weekly-genai 09-24 the Deception Benchmark.
• Reason for dropping it: each is a single piece and its momentum has not been confirmed (Hamel's FAQ is a candidate for a close read in a later run).

18. 🧩 Google ADK and Antigravity updates

• Dropped / impact 1 / potential 1 / anchor undated / track: GenAI engineering.
• Feed items in this cluster (google-developers-blog, undated): local model support in the Antigravity SDK (Gemma 4 26B A4B on LiteRT), ADK for Kotlin 1.0, evaluation of live / voice agents in ADK, zero-trust agents with ADK, and 4 patterns from the AI Agents Challenge.
• Reason for dropping it: the feed carries no dates, so the anchor cannot be fixed (check the page dates next time).

19. 📎 Other one-off items

• Dropped / impact 1 / potential 0 to 1.
• aws-whats-new 09-23 Amazon CloudWatch Omni GA (the primary source does not make clear what it offers for agents).
• aws-whats-new 09-18 credential testing support in AWS Continuum (a penetration-testing agent).
• aws-whats-new 09-22 agent-to-agent collaboration (A2A) in Amazon Connect Customer.
• microsoft-foundry-blog 09-15 Foundry Dev Pack.
• hatena-llm 09-25 the AI-SQL engine "Quail" (Shreya Shankar, jointly with Modal).
• publickey 09-16 macOS virtual environments for Devin, 09-13 Tailscale Aperture GA.
• hatena-claude-code 09-13 Spotify cutting Claude Code token consumption by 90%, 09-17 Minorun-style (みのるん式) Claude Code material, 09-19 Claude Mods (a provisional spec in Issue #91870), 09-22 Plugins in Claude Code.
• zenn-trending 09-23 a big clean-up of MEMORY.md, 09-25 a self-service telemetry platform built on OpenTelemetry.
• hatena-ai-agent 09-26 a round-up of web search options for local LLMs.
• pragmatic-engineer 09-17 AI Skills with Matt Pocock, 09-22 Windows and AI, 09-23 Design Engineering with Maggie Appleton.
• hn-llm 09-17 How to Write with an LLM (767 pt), 09-24 Best LLM for every budget.

20. 🚫 Out of scope

• anthropic-news 09-17 Life Sciences Verification Program, 09-18 embedded evaluation with Accenture, 09-23 enzyme discovery with Claude (scientific applications and partnerships).
• openai-news 09-21 to 09-25 customer stories (V7, Higgsfield, Parallel, Airbnb, Harvey, invideo, Ringg, Proaction), Academy, the policy on standardisation and third-party evaluation, ChatGPT Ads, and remarks at the UN Security Council.
• ainews 09-24 Meta Connect 2026 (devices), 09-25 the future of Latent Space.
• google-developers-blog reproducing OLMo 3 in MaxText (training infrastructure), autonomous post-training with Tunix, Colab, an essay on the Go language, edge AI on Raspberry Pi.
• publickey 09-13 Shopify's acquisition of Tailwind, 09-15 Java 27 and Java security patches.
• hn-claude-code 09-13 Houthi use of Claude Code (reported by Anthropic), martin-fowler 09-16 / 09-24 Fragments (commentary on AI-driven hacking and AI risk), 09-15 / 09-24 posts on presentations and feedback.
• hatena-llm 09-25 Nemotron 3 Diarization (speech), hn-llm 09-14 OpenArch, 09-15 the share of LLM-generated content at F-Droid, 09-20 Pirate Face, 09-22 LLM Ass Bench.
• gcp-vertex-ai-release-notes 09-13 to 09-21 maintenance releases of Agent Platform Workbench.
• aws-whats-new 09-15 / 09-16 shift bidding and Connect Talent in Amazon Connect, and instance priority lists in SageMaker.
• armin-ronacher 09-14 Interpreting Pangram (an essay on an AI detector).
• laiso-newsletter 09-13 a look back at running the newsletter in August.