Claude Managed Agents: Cap Research Spend and Record Advisor Intervention
Claude Managed Agents can pause at a hard session budget and consult an advisor model. Record both controls without treating them as evidence of completion or accuracy.
Direct answer: Claude Managed Agents can now stop a session at a hard spend budget and let the primary thread consult an advisor model. Use both controls as part of a research-run manifest: the budget limits new model requests, while the advisor record shows when another model influenced the plan. Neither control verifies the final claim.
This guide explains the documented controls and a publication workflow. It does not report an original benchmark of Managed Agents, completion quality or advisor value.
Set the budget from the reader job
Define the task boundary, required sources, expected artifact and stopping conditions before choosing a number. A narrow release-note comparison should not receive the same budget as a patent-family review or multi-market citation study.
Anthropic says the cap is priced at public list rates. When the session reaches it, the session pauses with a budget_reached stop reason instead of starting new model requests. Changing or removing the budget can resume the work. That behavior makes unfinished state part of the method.
| Field | Why it matters | Completion evidence |
|---|---|---|
| Budget | Bounds model spend | Configured amount and pricing date |
| Stop reason | Explains why work ended | Completed, budget reached or failed |
| Required sources | Prevents shallow completion | Source ledger status |
| Resume decision | Controls extra spend | Named approver and reason |
Treat the advisor as an intervention
Anthropic documents an advisor as a model at least as capable as the agent’s own that the primary thread can consult mid-turn for strategic guidance. Record the primary model, advisor model, request time, reason for consultation, returned guidance, action taken and whether a human reviewed the change.
An advisor can help with planning, contradiction checks or task decomposition. It is not an independent source, a fact-checking certificate or proof that two models agree. When both models share training patterns or incomplete context, the consultation can reinforce the same error.
Keep source selection and claim verification outside the advisor score. A primary document that contradicts both models wins.
Preserve the unfinished work
When the budget pauses a session, save completed steps, open claims, failed tools, missing sources, partial files and the next safe action. Do not publish a fluent partial draft merely because the session produced one before stopping.
- Check whether every material claim maps to a source.
- Mark missing fields and unresolved contradictions.
- Estimate the additional task, not only the additional tokens.
- Decide whether to resume, narrow the article or hand off to a person.
- Create a new run version when the budget, model, tools or scope changes.
A deployment-level budget applies to each session it starts, according to Anthropic’s release notes. Model the possible number of sessions so a per-session cap is not mistaken for an organization-wide ceiling.
Measure cost and completion separately
Report spend, elapsed time, source coverage, completed reader tasks, verified claims, rejected claims, human review time and rework separately. A cheaper run can be incomplete; a more expensive run can still be wrong.
Define success before seeing the output. For a source-led article, success may require every current product fact to resolve to official documentation, every uncertainty to remain visible, and the final package to pass a skeptical human review.
Use the source-bounded AI brief workflow before the run and the publication gate before release.
Ask a question or join the discussion