Resource Budgeting

Core rule

Allocate scarce resources before the workflow discovers the limit by crashing.

Name each relevant budget and reserve headroom for completion/cleanup: context plus output tokens, disk plus staging copies, API quota plus retries, memory plus peak build usage, process/MCP slots, and wall-clock timeout.

Estimate the dominant cost driver and monitor it during long work. Batch operations where it reduces repeated overhead, but cap concurrency where parallelism multiplies memory, rate-limit, or failure pressure.

When near a limit, degrade deliberately: prune context, checkpoint state, reduce fan-out, or switch to a supported cheaper path. Never spend the final capacity on optional work and then lack enough budget to verify or package the result.