Claude's API adds message threads, three Claude models start carrying a text watermark, Kimi K3 gets in-country inference in India on Bedrock
No new models today; the thread is running agents as production systems: who holds the conversation state, whether output is marked, whose identity an agent calls with, and treating skills and instructions as dependencies to update and review.
Enabled with the message-threads-2026-08-12 beta header: the first request creates a thread with the full conversation, system prompt and tools; later requests send only the new messages and the previous response id
Threads are best-effort: they expire 5 minutes after the last request and are deleted no more than 1 hour after last use; a continue request on a lost thread returns 404, and the full conversation must be replayed
system and tools are create-only, while fields such as model, thinking, tool_choice and mcp_servers must match on every turn or the request fails with thread_fingerprint_mismatch; a thread does not cache prompts by itself
Available on the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry; not ZDR or HIPAA eligible
Builder's takePandaClaws' background jobs are exactly this kind of agent loop, resending a long history every turn. I'd adopt threads through the SDK helper, which handles the replay, keep our own copy of the conversation as the source of truth, and add a top-level cache_control, because threads alone don't lower token cost. Anything covered by a ZDR or HIPAA agreement stays off it.
From November 3, text generated by Claude Mythos 5, Claude Opus 4.7 and Claude Sonnet 4.6 carries Anthropic's text watermark on the platforms listed in Anthropic's help article on how Claude marks AI-generated content
On Sonnet 4.6, text from a request that sets temperature, top_p or top_k to a non-default value might not carry it; if you rely on the watermark, Anthropic says to omit all three, including values a framework or client library sets for you
The linked documentation says the watermark adds no tokens or hidden characters, does not change meaning, quality or readability, and carries no information about the user or organization
Builder's takeNo code changes needed, but two checks. If a feature promises to disclose or detect AI-written text, confirm which models and platforms are covered instead of assuming all Claude output is marked. And if you pass sampling parameters to Sonnet 4.6 out of habit, decide whether you actually want them. In something like AI Interview, where people will ask whether an answer was AI-written, a watermark that depends on the model and the settings is a signal, not proof.
Kimi K3 is offered through cross-Region inference profiles: global.moonshotai.kimi-k3 is recommended for most use cases as the lowest-cost and most broadly available; in.moonshotai.kimi-k3 keeps processing within Mumbai and Hyderabad
AWS recommends the OpenAI-compatible APIs for this model, and its example connects through the openai Python library
The walkthrough reads a scanned Hindi document, extracts structured data and answers follow-up questions in Hinglish, with no separate OCR or translation service
Customer Ultrahuman says its assistant handles more than 14,000 conversations a day on Kimi K3, and that prompt caching cut its LLM costs by 80% and made multi-turn chats 85% faster
Builder's takeThe pattern matters more than India: open-weight models are getting region-pinned profiles on the big clouds, which is the first thing regulated customers ask for. If you serve markets with residency rules, set the geographic profile explicitly in config instead of relying on the global default, and check that your SCPs allow both destination Regions, or calls will fail in ways that look like model errors.
Dependabot opens pull requests to update supported git dependencies declared in apm.yml and regenerates apm.lock.yaml when it is present
Covers version updates only, not security updates
Set up by adding an apm entry to .github/dependabot.yml with the directory that contains apm.yml and an update schedule
Generally available on github.com for all plans; coming to GitHub Enterprise Server 3.24
Builder's takeSkills and instructions now have a lockfile, so treat their updates like code: read the diff in each PR, because a changed instruction can change what an agent does with your repo's credentials. With no security updates in scope, I'd still restrict skill sources to publishers we trust, the same allowlist idea we already apply to editor extensions.
Select GitHub Copilot in Air's agent picker and use an existing Copilot subscription; track several agent sessions in parallel and review changes next to the code
Cloud tasks run through Air Teams in a Docker container on a fresh virtual machine, with all traffic through a proxy and no state carried between tasks
You cannot choose whether a task uses a personal or an organization Copilot plan, Air's analytics do not include Copilot usage, and Air admins can disable Copilot but cannot set Copilot spending limits in Air
Builder's takeACP is turning the IDE into a picker for whichever agent you already pay for. The catch is in the fine print: no plan selection and no spending limits inside Air. If a team adopts it, set Copilot budgets on the GitHub side before turning on cloud tasks, and expect Air's dashboards to miss that spend.
The language model handles workflow orchestration, database navigation and reasoning, while deterministic tools built on RDKit do the molecular calculations, with connectors to ChEMBL and PubChem
In a walkthrough based on crizotinib and lorlatinib, the agent retrieves thousands of compounds, runs matched molecular pair analysis, builds potency models validated on a held-out set and proposes rigidified macrocycles with synthetic routes
DeepMind says such a workflow typically takes multiple days per iteration, and that proposed designs are testable hypotheses that must be synthesized and validated experimentally
Available today by installing the Science plugin in the Antigravity app
Builder's takeThe transferable part is the split: let the model plan and pick tools, and push anything that must be exact into deterministic code. That's how I'd build extraction in AI Cloud Drive too: the model decides what to pull from a file, and a parser or validator produces the final fields. Packaging the know-how as a plugin also means domain rules ship separately from the model.
Four difficulty levels map to kinds of variable assignment: immutable values, reassignment by value, self-referential updates (x = x + 5) and assignment by reference (x = y, swaps)
All four frontier models did well on the two easier levels, but on the two harder ones accuracy fell sharply as inputs grew, to 13% or lower on the longest general finite problems
Trained from scratch at 275M parameters on sequences of up to 256 assignments and tested on up to 64,000, Gated DeltaNet and hybrid models generalized almost perfectly on three levels; transformers degraded with length and never learned the hardest level
Format matters: some small models scored near the ceiling on RULER-style problems but near chance on the same problems framed as spreadsheet edits
Builder's takeThis is the failure behind agents that lose the plot after many edits: working out the current value after a chain of updates is hard even for frontier models. In long agent runs I wouldn't ask the model to reconstruct state from the transcript; keep the current state in a structured object the harness updates and feed it back each turn. Test with your own data formats too, not just needle-in-a-haystack scores.
An agent gets a checkout of a real open-source repository and a short instruction, and must find and fix a security weakness without breaking existing functionality; scoring is fully programmatic, and a held-out set is kept private
Of 11 models, Gemini 4 Argon leads with 67.8% pass@1 and 81.0% pass@4, ahead of Claude Opus 5.5 at 67.3% and 78.0%; even the leader fails 19% of tasks on every attempt
On cost, DeepSeek V4.1 Flash (1017), Claude Opus 5.5 and Gemini 4 Argon form the Pareto frontier: 57.0% pass@1 at $0.04 per rollout, about ten points more for Opus at $1.83, and Gemini slightly ahead at $3.92
95 tasks are solved by at least one model; the public set, grader and harness are open, and tasks are tagged as RL environments for training defensive agents
Builder's takeThe cost column is the useful one: a cheap open model gets most of the way, which suggests a two-tier security pass in CI, with the cheap model sweeping every PR and a frontier model on what it flags. Whatever the model, keep the benchmark's own bar for agent fixes: existing tests must still pass, and a person signs off on security patches.
Without identity propagation, agent calls reach downstream services as the application rather than the user, which breaks audit trails and bypasses user-level access controls
The agent authenticates to the gateway with a JWT scoped to its own Microsoft Entra ID app registration; the gateway exchanges it through the OBO flow for a token scoped to the ServiceNow connector, with the user's identity preserved
ServiceNow maps the token's preferred_username claim to an existing user and applies that user's permissions; the exchange is configured on the gateway target and runs on each request
AWS says the pattern applies to other REST APIs protected by OAuth
Builder's takeThis should be the default for any agent that touches business systems: the agent acts as the user, not as a super-account. For AI Cloud Drive's agent features I'd do the same, exchanging the user's token at the tool gateway and letting the storage layer enforce its own permissions, rather than filtering results inside the agent.
Researched and drafted with AI assistance; editorial standards and views set by Darius.
Every Monday, the week in AI in one email
Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.
Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →