Budget policies for overnight background agents: preventing infinite retry loops

Some-Beat5994 · reddit · 2026-08-26

A developer asks for practical budget policies for background agents that can retry overnight. The primary risk is not a single expensive request, but a small error triggering repeated tool calls or retries over several hours. The author seeks strategies to limit spend while allowing recovery from transient failures. Discussion points include per-task token budgets, retry ceilings, time-based escalation, and separate model paths for planning, execution, and verification. The author also asks how to distinguish recoverable tool failures from tasks requiring human intervention.

Original post →

More from coding & agent

coding & agent channel →