Anthropic's Managed Agents add spend caps, an advisor, GitHub skills
The server-hosted agent runtime can now hard-cap a session's spend, pin inference to a region, call a stronger model mid-run, and load skills from a repo.
Copy markdown
A hard spend cap, per session
Managed Agents sessions now take a budget priced at public list rates; when it's hit the session stops with a `budget_reached` stop reason instead of firing new requests. Deployments can apply the same cap to every session they spawn — so a runaway agent loop can't quietly drain your balance.
An advisor model, on call mid-turn
Add a second model — as capable or more capable than the agent's own — as a `{"type": "advisor"}` entry in the multiagent roster, and the primary thread can consult it for strategy without handing off the turn. Run a cheap worker and pull in a smarter brain only when it's stuck.
Pin inference to a region
A new `inference_geo` field on the model object sets where inference actually runs, per agent or overridden per session. Data-residency rules that used to rule out Managed Agents — EU and others — no longer have to.
Skills load straight from a repo
Mount a GitHub repo and any skills in its root `.claude/skills` directory are discovered automatically at session start. Version your agent's skills in git next to your code — no separate upload step.