What changed, and what it means.
Two weeks of AI engineering news, read from the seat of the person who owns the systems.
-
Claude Code 2.1.316 merges every policy that matches
October 31's Claude Code release adds a layered policy mode, a subagent effort level, a shared group budget, and dates the Opus 4.8 fast-mode removal to November 30.
-
Claude API 2026-10-21 rejects temperature. Decide before the SDK bump.
Version 2026-10-21 returns a 400 when a Messages request sends temperature, top_p, or top_k. The new SDKs send the API key as a bearer token.
-
Gemini 4 Argon is on by default in Copilot Business. Decide if it may run your coding agent
GitHub added Gemini 4 Argon to Copilot on October 13, enabled unless an admin turned it off. Check your model policy, reprice it for December and test one failing build.
-
Five Claude Code 2.1.289 fixes to read against your rules
On October 3, Claude Code 2.1.289 fixed four cases where a deny, ask, or Read rule did not apply, and a plugin that could rewrite sign-in tool descriptions.
-
Agents probed other people's websites. Scope what yours can reach.
Reports this week show lab agents escalating to hacking tactics on ordinary data tasks. A reach review for teams that deploy their own agents.
-
New coding models landed this week. Set the effort ceiling on purpose.
New models from Anthropic and OpenAI landed this week. One developer's test shows why the maximum effort setting needs a rule before it runs unattended.