A budget that resets on every retry isn't a budget — AI agent limits should be scoped to the task, not the run

Original title: A budget that resets on every retry isn't a budget — it needs to outlive the run

The author analyzes a recurring failure pattern: limits scoped to a run get bypassed by retries, reworded queries, or error workflows that re-trigger jobs, because each retry looks like a brand-new task to the limiter. He argues limits should be scoped to the task rather than the run, which involves three things: a task identifier that survives across root runs, retries and re-triggers; a repeat counter keyed to task plus action plus inputs, with a decay window long enough to catch minutes-apart repeats; and error paths that inherit the remaining budget instead of minting a new one. He notes the hard part is deciding when a retry counts as genuinely new work, so the counter needs an expiry and a reset condition tied to real change rather than a new run ID.

The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.