Claude's Managed Agents get hard spending caps
By Steven Van ·
A session now pauses at its budget instead of terminating, and resumes where it left off once you raise the cap.
Anthropic shipped four updates to Claude Managed Agents this week, and the headline feature answers the question everyone running autonomous agents eventually asks: how do I stop this thing from quietly burning through my budget? You can now set a hard spending cap on any Managed Agents session. When the session hits the limit, it pauses with a budget_reached event instead of terminating, and raising the budget picks the work back up exactly where it left off.
A spend cap that pauses the agent instead of killing it
Managed Agents is Anthropic's hosted agent platform: you define an agent once, and Anthropic runs the loop and provisions a sandboxed workspace for every session. That autonomy is the appeal, but it is also the risk. A long research run or an overnight coding session can rack up model tokens, web searches, and compute time while nobody is watching.
Session budgets fix that with a design choice worth appreciating. The budget is a hard dollar cap you set when you create the session, and the platform checks spend before every model request. Once the cap is reached, the session goes idle with a budget_reached stop reason. Nothing is lost: the conversation history and the sandbox are preserved. If the paused work turns out to be worth finishing, you change the budget to a higher number, or remove it entirely, and the session resumes automatically. If not, you archive it and walk away knowing exactly what it cost.
What counts toward the cap
The budget tracks the session's consumption priced at Anthropic's public list rates: model tokens at each model's list price, web searches at 10 dollars per 1,000, and session running time at 8 cents per hour. The cap is denominated in US dollars and specified in cents, so a value of 2500 means a 25 dollar ceiling. Anthropic notes that enforcement is a gate on new work rather than a mid-request kill switch, so the final figure can run slightly past the cap by at most one in-flight request per running thread.
Scheduled deployments, the feature that fires agent sessions on a cron schedule, accept the same budget object and stamp it onto every session they start. That combination is the practical win: a nightly report generator or a weekly content pipeline now has a per-run cost ceiling baked in, no monitoring script required.
The other three updates: advisors, data residency, and repo skills
The budget feature led the announcement, but the same week's release notes list three more Managed Agents changes.
First, sessions can now have an advisor: a model at least as capable as the agent's own that the session consults mid-turn for strategic guidance, like planning an approach or reviewing work before finishing. You configure it as an advisor entry in the agent's multiagent roster and name the model to consult, so a cheaper agent can borrow a smarter model's judgment only at the moments that matter.
Second, you can now pin where model inference runs by setting inference_geo inside the agent's model configuration, or override it for a single session. For teams with data residency requirements, that turns a compliance conversation into a config field.
Third, sessions that mount a GitHub repository now automatically discover skills stored in the repo's root .claude/skills directory at session start. Your team's playbooks and workflows live in version control next to the code, and every agent session picks them up without a separate upload step.
Why predictable agent spend matters for creators
Agents are moving from demos to production workflows: research assistants that compile briefs, pipelines that draft and format content, bots that triage and summarize. The blocker for a lot of small teams has not been capability, it has been the fear of an unbounded bill. A cap that pauses gracefully, keeps the work intact, and resumes on your signal makes it much easier to hand an agent a real job and let it run. Combined with scheduled deployments, you can now put a fixed price on every automated run before it starts.
