Managed Agents budget_reached vs Sume generation_spend_cap_usd

They are two separate ceilings. A Claude Managed Agents session budget caps the session's spend at public list rates and pauses it with budget_reached; Sume's generation_spend_cap_usd caps generation spend on a Sume Agent Completion in dollars. Neither is documented as reading the other.

The Anthropic fact is from its release notes (entry dated 2026-08-07); Sume facts from Agent Completions, read 2026-10-01.

What does the session budget do?

Per the release notes it is a hard cap on a session's spend, priced at public list rates. A session that reaches it pauses with the budget_reached stop reason instead of starting new model requests; changing or removing the budget resumes it. Deployments accept the same budget and apply it to each session they start.

What does generation_spend_cap_usd do?

It has no default: omit it and the request fails with 400 invalid_request. The docs call it the substitute for the interactive spend-approval prompt, which a backend caller does not have, and say to set it to the most you will spend on a single run. The receipt echoes it as generation_spend_cap_usd_micros under usage.

Which one stops a run first?

Whichever limit the run reaches first, and the two are measured differently (list-rate session spend versus Sume generation spend), so they are not directly comparable. A run that spends mostly on model tokens meets the session budget; one that spends on generations meets the Sume cap. Set both deliberately, and size the Sume cap from the API pricing page as the docs suggest.

Is there a tighter bound inside a script?

Yes for MCP script_run: it is bounded by timeout_seconds, max_calls and max_paid_calls. Related reading: paid MCP calls under auto permissions.