Your AI subagent's report is all you'll ever see. Make it hand in receipts

If you use AI agents that spawn helper agents (audits, research, big refactors), here's the mechanical fact that changed our setup: the parent never sees the work. The helper runs in its own context; when it finishes, one final message comes back. The forty tool calls behind it — the searches that found nothing, the test that failed twice before passing, the edge case quietly ruled out of scope — stay behind. What you receive is a summary, compressed by the thing being evaluated.

That's not a bug. Isolation is why delegation is affordable at all. But it means every "Done, all tests pass" is testimony, and nothing in the loop charges the helper for optimism. A caveat is a paragraph; "done" is one word. Guess which survives compression.

Ours got caught the memorable way: an audit subagent confidently reported that a sourced claim had no source behind it. The source existed. We found the exact sentence in the original docs and ended the argument in minutes.

What we require now, roughly in order of payoff: evidence formats written into the dispatch prompt (every finding arrives with a verbatim quote and a file path), artifacts over adjectives (test output and re-runnable commands instead of "thoroughly checked"), and spot-checks scaled to blast radius — anything we'll act on gets one finding re-derived end to end. The re-check is deliberately cheap: when a claim comes holding a quote, verifying it is a text search.

Limits, honestly: we have a handful of caught errors, not a measured rate, and most reports are fine. But "most" is doing a lot of work in that sentence when you're about to build on one.

What does your dispatch prompt require the helper to hand back?