Anthropic donates Agent Skills to the AAIF, Claude arrives on Google Cloud as an Anthropic-managed API, OpenAI's Decisions API goes GA
Today's thread is agent parts turning into public infrastructure: the skills standard goes to a neutral foundation, the same API lands on another cloud's bill, and lightweight decisions get their own endpoint and price. On the training side, what an agent leaves behind is starting to be used directly as the reward.
A skill is a folder with a SKILL.md file that teaches an agent a task, plus any scripts or reference material it needs. Since it was published as an open standard in December 2025, more than 70 platforms have integrated it, including ChatGPT, Cursor, GitHub Copilot and VS Code.
The recipient is the Agentic AI Foundation (AAIF), which Anthropic co-founded with Block and OpenAI and which already holds MCP. Anthropic is handing over the Agent Skills name and logo, the GitHub organization and the agentskills.io domain; using or contributing to the standard doesn't require joining the AAIF.
Engineers from Anthropic, Google, OpenAI and other companies maintain the standard as individuals, with David Soria Parra and Mahesh Murag as Lead Maintainers.
The proposed first roadmap covers versioned releases, a formal proposal process, and security guidance for people who write and install skills.
Builder's takeI'd write the team's internal prompts and runbooks to the SKILL.md spec rather than any one vendor's format: with the standard in a neutral foundation, the same skills should move between Claude Code, Codex and Copilot as they are. PandaClaws' content pipeline is already a step-by-step procedure, which is exactly what a skill is for. Treat installing someone else's skill like adding a dependency; until the security guidance lands, only install skills whose scripts you've read.
It is the Claude API operated by Anthropic, reached from a Google Cloud project: Google Cloud IAM controls access and Google Cloud Marketplace handles billing. Fast mode for Claude Opus 5.5 is available too.
Setup creates a new Anthropic organization for the Google Cloud project, separate from any organization your company already has, Claude Enterprise included. Existing discounts need a call to your account rep before you start, and can't be applied to past usage.
There are three predefined IAM roles: claude.viewer is read-only, claude.user can send messages and manage batches, files, skills and Managed Agents resources, and claude.admin adds organization-level permissions and is the only role that can open the Claude Console.
Roles are granted on the project and apply to every workspace in it; when teams need separate access, the docs recommend a separate Google Cloud project per team, each with its own Anthropic organization.
Builder's takePut this next to the 10-22 item on OpenAI models coming to Google Cloud: within a week Google Cloud has both vendors' APIs on its shelf. If your team has committed GCP spend, you can now move Claude onto that bill too, but this is a new organization: workspaces, skills and usage history in your existing org don't come along, so make an inventory before migrating.
The Decisions API (v1/decisions) is generally available with gpt-6-luna. When the beta shipped on October 6, OpenAI said it turns text and images into typed answers 10x faster than the Responses API.
Each request now supports up to 400 questions, up from 200 in the beta.
Shared input can hit the prompt cache across requests for cache discounts, individual question instructions can include images, and structured data can be passed directly with input_json.
Builder's takeThe docs list uncached input on gpt-6-luna at $0.10 per million tokens with output not charged, which is cheap for read-this-then-answer-a-list-of-questions work. Tagging a résumé with a dozen labels in AI Interview has exactly that shape: put the résumé in shared input to hit the cache, make each label a question, and compare cost and latency against today's general-purpose call.
Fixed: on macOS and Linux, sandboxed Bash commands ran the shell's startup files outside the sandbox before each command.
Fixed: sandboxed commands and auto-accepted edits could modify the MCP config file passed with --mcp-config, and in some commands with a subshell an Edit(path) deny rule didn't block a Bash redirect to that path.
Fixed: on Linux, sandbox write denies lifted when a background command outlived the session; on Windows, PowerShell permission checks auto-approved some Set-Location or cd commands with extra arguments.
Changed: sed -i now gets the standard Bash prompt and runs as written instead of a simulated file-edit preview, and is no longer restored by /rewind; claude ultrareview now requires workspace trust.
Builder's takeIf you run Claude Code unattended in CI or scheduled jobs, upgrade today. This batch is also a reminder that the sandbox and permission rules are one layer; the real boundary is still the container and least-privilege credentials. And sed -i can no longer be undone with /rewind, so warn anyone who leans on /rewind as a safety net.
Training sessions extend SageMaker AI serverless model customization: you write the training loop, custom loss and reward signals, and SageMaker AI handles GPU allocation, distributed execution, fault recovery and checkpoint storage.
SFT, DPO, RLVR, RLAIF and multi-turn RL are supported, with LoRA or full fine-tuning. The reward function runs in your own environment, where it also connects the agent to its tools or simulators.
You pay for the tokens SageMaker AI processes; the meter stops while your reward function scores a batch, and there's no capacity to reserve.
A companion cookbook ships model-customization agent skills that work in SageMaker Studio, Kiro, Claude Code, Codex and GitHub Copilot.
Builder's takeNot paying while your reward function scores matters a lot for multi-turn agent training: a slow evaluation environment no longer burns GPU money. I'd take an end-state reward like the ThinkingBox one in the next item, run a few dozen steps on a small open model to see the curve, and only then decide whether a fixed task deserves its own specialized model.
Each of ThinkingBox-Bench's 507 tasks starts a real backend, gives the agent MCP tools and a simulated user with private context, and at the end runs executable checks on the final database state: whether the records are right, not which tools were called.
The team built a training pool from similar tasks outside the benchmark, choosing ones neither too hard nor too easy for the current model, based on success rates estimated once from the base policy.
After 50 GRPO steps, Qwen3.8-27B's average success rose from 51.70% to 60.78%, and tasks passed on all 20 attempts rose from 38 to 107, with a small decline in at-least-once coverage.
On a hotel downgrade task, the base model changed the booking without required manager approval in 15 of 20 attempts; after training it preserved booking and payment state in all 20.
Builder's takeEven without training, use this idea for evaluation first: when an agent organizes or moves files in AI Cloud Drive, acceptance should check what the folders and permissions end up as, not whether the agent says it's done. Tasks passed 20 out of 20 going from 38 to 107 tracks the production experience better than the average, and that metric is worth copying.
The global default policy announced in September begins rolling out enforcement today for Copilot Business and Copilot Enterprise, completing by October 26, so it may take effect at different times for different enterprises and organizations.
Generally available features under Features & clients and MCP servers in Copilot that are still Unconfigured will be set to the chosen global default of Enabled, Disabled or Let organizations decide. It's a one-time change; later changes to the default won't touch settings already in place.
Features you've explicitly enabled or disabled stay as they are, and preview features remain opt-in.
Builder's takeI'd check the default under AI Controls before the 26th, especially the MCP servers section: once it's set to Enabled in this one-time pass, changing the default later won't take it back. Better to start with Let organizations decide and open things up one by one.
Metrics include the share of users who are agent-first or multi-agent, and how the pull request merge rate of agentic users compares with nonagentic users, ranked against the peer median over a 28-day period.
Peer groups are set by company size (such as 1–999, 1,000–4,999 or 5,000+ employees) and one of ten industries; when an industry group is too small it widens to all industries for that size, and the dashboard labels the group it used.
It's opt-in and off by default; opting out stops comparisons immediately and excludes your data from the next daily update. Aggregated benchmarks carry no identifying data, are deleted after 30 days, and aren't in the usage metrics API or exports.
Builder's takeAs someone who runs a team, I'd watch the merge rate, not the agent-first share: the share is easy to inflate under a target, while merge rate shows the output was accepted. Don't drop the industry median straight into OKRs; use it to find gaps, then work out whether the cause is tooling or process.
The upgraded Generative engine produces more natural, more accurate speech with low-latency streaming, and adds 14 voices across 9 locales, including generative voices in Hindi and Finnish for the first time.
It's a drop-in upgrade for existing generative voices: voice names and API calls are unchanged, so existing applications keep working.
Pricing is unchanged with no new SKU, and synthesized static prompts can be cached and replayed at no extra cost.
Builder's takeA drop-in upgrade means your production voice has already changed under you. For a voice product like AI Interview, I'd re-listen to a few fixed scripts that read out IDs, names and jargon, and keep them as an audio baseline; fixed prompts like the opening line can switch to cached replay, which is cheaper and more consistent.
Researched and drafted with AI assistance; editorial standards and views set by Darius.