When you left a Claude agent running, there was no way to know how far the cost would go. This week, Anthropic released four updates to Claude Managed Agents at once, including a session budget feature.
It Stops When the Limit Is Hit: budget_reached
The first feature out is the budget setting. If you put a budget on a session, the moment it reaches that limit the session stops automatically with a budget_reached event. To keep it going, raise the budget and resume. Before, if you left a long-running automated agent on, you only found out what it cost after it ran. Now you can set a ceiling in advance.
The Other 3 Updates That Landed in One Week
| Update | What it does |
|---|---|
| Inference region pinning | Set model.inference_geo to us or global. Global uses standard pricing wherever there is capacity; us pins the region at 1.1x pricing |
| Automatic skill loading | Automatically loads .claude/skills/ from the connected repository when a Managed Agent session starts |
| Advisor | A working agent can call a stronger model mid-session for a second opinion; added with a single line of configuration |
The advisor feature is especially practical. You can now build a setup that runs repetitive work on a cheap model and calls a higher-tier model only at points where the judgment is unclear, with one line of code.
The Longer It Runs, the Sooner You Lock the Budget
1. As long-running automation grows, cost control came first. When an agent keeps working for hours on its own judgment, it is hard to predict where the cost will stop unless a person watches every moment. budget_reached is a device that draws a ceiling on that uncertainty.
2. Region pinning addresses regulation and data policy. The ability to tie inference to a specific region with inference_geo removes one condition that comes up when organizations with data governance requirements adopt cloud agents.
3. The advisor is an attempt to get cost and quality at the same time. It means a structure that does not leave every judgment to an expensive model, and calls it only when needed, is now a standard setting.
Already Used Up the Weekly Limit?
Comments on the announcement thread also included complaints such as "usage seems to run out strangely fast over the past few days" and "give us a weekly reset." The budget cap helps with predicting spend, but some noted that what users really want is a higher limit itself rather than tools to watch the limit closely. There is still a gap between the feature announcement and how users feel.
A Cost Ceiling First, Finer Control Next
All four updates move in the direction of making it possible to keep agents on longer and more safely. Budget, region, skills, and advisor read as the conditions that held back putting automation into production being filled in one by one.
Related: Official ClaudeDevs X thread