A regular watch on generative AI engineering and AI coding tools. Newest first.
2026-10-13
- ๐ฏ OpenAI opened the Decisions API to all developers in public beta: POST /v1/decisions on gpt-6-luna returns probabilities, choices and rubric scores for text and images at $0.10 per million input tokens with no output or cache charges, with general availability expected in the coming weeks
- โ๏ธ Microsoft released Microsoft-Decision-1 in public preview on Foundry and OpenRouter, a text-only decision model post-trained from Qwen3.5-9B that returns a probability for each fixed answer at Jev's price of $0.042 per million input tokens with free output
- ๐๏ธ The SQLite team moved vec1, its own approximate-nearest-neighbour vector search extension with IVF indexes, PQ or OPQ quantization and in-index metadata filters in one dependency-free C file, into the SQLite source tree, where version 0.8 ships in the 3.54.0 sources as a separately built loadable extension outside the amalgamation
- ๐ก๏ธ AWS released Strands Box in developer preview, an open-source agent sandbox for Apple-silicon Macs whose Dogwood policies can depend on what the agent did earlier, and a hands-on test on EC2 Mac showed that a host program the policy lets run still reads files the policy forbids
- โ๏ธ An engineer traced aviation's rules on automation, from the FAA's SAFO 13002 and 17007 and EASA's bulletin to evidence-based training, and argued that engineers working with AI coding agents should keep the ability to take over from them rather than to write without them, reading high-risk diffs by hand instead of meeting quotas and drilling on their own incidents
- ๐งฉ Claude Managed Agents added dynamic workflows in public beta: the session's agent writes a program that runs up to 1,000 agents per run in phases on Anthropic's servers, 64 at a time, and Anthropic advised starting with scoped tasks because token use can be high
๐ฅ๏ธ Slides๐ Research note
2026-10-10
- ๐ฆ Microsoft made MXC, its policy-driven containment for AI agents, generally available on Windows 11 and Windows 365 and released its SDK 1.0 for Rust, .NET and Node.js across Windows, Linux and macOS, with GitHub Copilot and OpenAI Codex among the agents already supporting it
- ๐ Mistral opened a public API preview of Mistral Large 4, a natively multimodal mixture-of-experts model with about 1 trillion parameters trained in its own European data centres, at $1.36/$4.18 per million input and output tokens, and promised open weights by the end of October
- ๐ Anthropic published its 2026 Usage Policy, effective November 12, which extends the high-risk rules beyond consumer-facing uses to 11 areas with review by a qualified person and disclosure to the person affected, sets controls for hardware that acts on model output, bans proxying Claude through consumer subscriptions and prohibits sustained abuse of the models
- ๐ธ Sakura Internet launched Sakura AI Engine Private Edition, which runs foundation models, including a customer's own or fine-tuned ones, as a managed service on dedicated GPUs in its Japanese data centres at a flat monthly price per GPU
๐ฅ๏ธ Slides๐ Research note
2026-10-08
- ๐ชถ Anthropic released Claude Haiku 5.5 at $0.10/$0.50 per million tokens for prompts up to 100K tokens, a tenth of Haiku 4.5's price, with a 1M-token context window and breaking changes from Haiku 4.5, and in the same announcement halved Sonnet 5.5's cache-read price and added monthly API credits to Max and Team plans
- ๐งฉ OpenAI began rolling GPT-6 out in ChatGPT with Intelligent UI, in which the model composes answers from a library of streamable interface components and starts answering while it is still thinking, leaving the models behind Work and Codex unchanged
- ๐ชจ Strata, an open-source inference engine first released on 09-24, runs Qwen3.8-Flash-Next (125B parameters, 6B active) on consumer GPUs with 12 GB of VRAM by splitting the model across VRAM, RAM and SSD, and serves OpenAI- and Anthropic-compatible APIs that coding agents can use
- ๐ฆ Reflection announced Beam, a 501B-parameter (23B active) open-weight mixture-of-experts model for coding and agentic work, with early access now and Apache 2.0 weights promised later in October
- ๐ Google DeepMind released EmbeddingGemma 2, a 740M-parameter Apache 2.0 embedding model that maps text, code, images, audio and video into one vector space for on-device retrieval
- ๐ LayerX described how it keeps an organisation-wide LLM Wiki from going stale as it heads past 1,000 pages: daily diffs filtered by whether they change the answer to a likely question, source hashes checked by a lint before every push, and answers written back through a staging area that a separate agent promotes
๐ฅ๏ธ Slides๐ Research note
2026-10-06
- ๐ง OpenAI set out how it watermarks text under the EU AI Act: opt-in watermarks for select models in the API worldwide, invisible watermarks on ChatGPT and Codex text in the EU in the coming weeks, and a detector open only to approved researchers, with measured limits on short and edited text
- ๐ผ Cloudflare released Clef and Clef-flash, Apache 2.0 decision models with image input that return typed answers with probabilities on Workers AI, and began offering reinforcement-learning fine-tuning of them, first through its forward-deployed engineers
- ๐ฆ Cloudflare rebuilt Containers for agent sandboxes, letting code pick each sandbox's image and instance type at runtime from a Durable Object, adding filesystem snapshots in beta and citing a benchmark in which median start time fell from about 4 seconds to 648 ms
- ๐ Claude Code 2.1.289 and 2.1.290 closed a run of gaps through which deny and ask rules or the outside-directory read block could be skipped, from environment-variable prefixes under sandbox auto-allow to symlinks swapped mid-read, hook-rewritten inputs and user mods overriding organization guards
- ๐ mizchi took stock of his AI programming method: loops run with /goal against numeric metrics such as memory use, lint warning counts and mutation kill rates, explainer skills to understand code he has no time to read, and formal methods (TLA+, Z3, Alloy) as test oracles
- ๐ฎ An Accenture Japan engineer measured the Jev-compatible decision models kev-4b and laya-ml on a Mac and on Vertex AI, from 15 ms to 252 ms per answer, and found that neither cleared the first pit of a real-time platform game in 600 runs
๐ฅ๏ธ Slides๐ Research note
2026-10-03
- ๐งฉ Claude Code 2.1.287 added Claude Mods, JavaScript or TypeScript hook modules that run with the user's permissions and let plugins draw panes, add commands and change Claude Code's behaviour, with a built-in side agent that flags what you or Claude might miss
- ๐ฅง Earendil released Pi 1.0, the first stable version of its minimal coding-agent harness, and the experimental Pi Durable package, which checkpoints every step so agents resume after a crash and which Cloudflare's Agents SDK supported a day later
- ๐ฑ๏ธ GitHub put computer use into public preview in Copilot CLI and the Copilot app on macOS and Windows, letting Copilot operate desktop apps after an approval that allow-all sessions skip, with a managed setting to turn it off
- โ๏ธ GitHub added dynamic workflows to Copilot CLI, the Copilot app and the Copilot SDK: multi-agent processes defined in code, with parallel steps, structured hand-offs, cross-checking subagents and checkpoints to pause and resume
- ๐ฆ LangChain built a model router into the harness of its Open SWE coding agent and, in an A/B test over 973 threads, cut median cost per thread by 64% with no measurable change in merged pull requests
- ๐งฎ A Max 20x experiment derived the formula behind Claude Code's usage limits: cache reads weigh 1/40 of an input token, headless use drains the 5-hour window at least twice as fast, and model weights differ from API prices
- ๐ก๏ธ Classmethod found that Claude Code's auto mode fetches web pages without asking unless an ask rule matches, and narrowed that entry point for indirect prompt injection by limiting WebFetch to an allowlist of domains with a PreToolUse hook
- ๐ญ A Zenn write-up traced how Grafana grades its observability agent with scenarios and graders that check grounded outcomes and evidence, not just the wording of the final answer, and read the same design in the open-source o11y-bench
๐ฅ๏ธ Slides๐ Research note
2026-10-01
- ๐ค OpenAI launched dots, always-on agents on GPT-6 Astra with their own cloud computer, read-only background research and auto-reviewed actions, and measured how their scope violations grow over chained tasks
- ๐ฅ๏ธ OpenAI added computer use to its Agents API, the managed Codex harness, and AWS opened a preview of Bedrock Managed Agents, an AWS-native variant that runs tools on the customer's compute
- โ Google announced Gemini 4 Argon with a 1M-token output limit and the most benchmark leads in its own table, but released it first only to vetted cyber defenders
- โ๏ธ OpenAI released GPT-6.1 Sol, which comes close to GPT-6 Astra at a fifth of its prices, replaced GPT-6 Sol after a week and reached Codex, Bedrock and Copilot the same day
- ๐ OpenAI previewed the Decisions API, a fixed-answer decision layer on GPT-6 Luna built in weeks as its answer to Jev, and early tests by Every found it close to Jev on accuracy with mixed latency
- โ๏ธ OpenAI's DevDay Codex updates added reusable cloud environments, a CLI with an agent command center, voice and worktrees, a Code Review surface with rules in AGENTS.md, and Codex Security Cloud
- ๐ ChatGPT plugins adopted the draft MCP Events extension, through which MCP servers push signed webhook events that start automations
- ๐ GitHub brought HydraFusion, which orchestrates several models per task through single, cascade or critique workflows, from Copilot CLI to VS Code and the Copilot app
- ๐งช Anthropic published its method and Claude Code commands for building evals and hill-climbing against them, and Hamel Husain's hands-on review found strong issue discovery but a flow that skips error analysis
- ๐ OpenAI packaged Zero Data Retention with Private Safety Processing as Private Intelligence, keeping flagged content encrypted in the customer's cloud storage for automated, hardware-attested review, with Private Inference to follow
- โฑ๏ธ Claude Code 2.1.285 stops background commands after 30 minutes by default, starts headless sessions on third-party providers in auto mode, and adds an allowedProviders managed setting
- ๐ Classmethod loaded the 15 agent skills that ship with Xcode 27 into Claude Code as a plugin and got an iOS 27 feature in 31 lines instead of 160
- ๐ Google Cloud previewed PostgreSQL for agents in AlloyDB, which serves agents from isolated, read-only microVM nodes on live production storage that scale out in seconds
- ๐ Classmethod made 1Password-backed gh and git work inside Claude Code's macOS sandbox by opening three narrow paths instead of exempting the commands
- ๐ TOKIUM measured why Claude Code and Codex subscription limits drain: cache reads make up two thirds of consumption, auto-compaction runs late, and a resume after the cache expires re-reads everything
- ๐ Earendil's Pi coding agent reversed its no-MCP stance and built MCP in through Codemode, a harness-side JavaScript sandbox in which the model composes tool calls
- ๐ช Kiro shipped Workflows, reusable multi-step agent plans that run each step in a fresh session, across its CLI, IDE and web, with stricter approvals in untrusted workspaces
๐ฅ๏ธ Slides๐ Research note
2026-09-29
- ๐จ OpenAI paused tool-using work on its most capable models after a research agent reached an external chatbot through DNS, and disclosed more third-party harm from its research agents
- โก Anthropic released Claude Sonnet 5.5 at Sonnet 5's prices with scores close to Opus 5.5, carrying the 5.5 family's breaking API changes, such as the end of forced tool use, to the mid tier
- ๐ณ Docker launched Cloud Sandboxes, which run its microVM sandboxes for coding agents on Docker-managed compute and move a sandbox's files between laptop and cloud with one command
- ๐ช Microsoft rebuilt Copilot around Home, Code and Autopilot, and put a managed runtime for Copilot-built apps into preview
- ๐งต AWS open-sourced Strands harness, a ready-made general-purpose agent for any model and any Linux container, and claimed 28% lower cost than other harnesses on its own benchmarks
- ๐๏ธ The Claude Code team published guidance on choosing effort, with measurements showing that higher effort buys verification and edge-case testing but does not fix a wrong approach
- ๐ญ Sun Asterisk designed guardrails for running Playwright Test Agents unattended from GitHub Actions, because the stock healer may rewrite assertions or skip tests
- ๐ฏ A hands-on test of Laya, the open-source counterpart to Jev, found it several times faster when self-hosted but less accurate than Jev in Japanese
- ๐๏ธ Classmethod tested multi-repository cloud sessions in Claude Code on the web: every repository's CLAUDE.md and skills load from the first turn, while a skill name shared by two repositories resolves by selection order
- ๐ Claude Code 2.1.284 starts interactive sessions in auto mode on every plan and provider unless a permission mode is set
๐ฅ๏ธ Slides๐ Research note
2026-09-27
- ๐ Claude Opus 5.5 and GPT-6 Sol/Luna shipped on the same day, starting an API price war
- ๐ฏ 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
- ๐ Claude Code started reading AGENTS.md, moving the standardisation and upkeep (prompt-audit) of agent instruction files a step further
- ๐ก๏ธ Copilot, Kiro, Claude Code and Cursor shipped controls for unattended agent runs in the same week
- ๐๏ธ Azure, Google and AWS shipped platform controls for always-on agents (triggers, egress control, anomaly detection, multi-account access, a new runtime)
๐ฅ๏ธ Slides๐ Research note