DEV Community

rambo
rambo

Posted on Edited on

How to Verify Your AI Agent Actually Completed the Task

#ai #agents #productivity #tutorial

How to Verify Your AI Agent Actually Completed the Task

Your agent says "done." Cool. Now prove it.

This is the part of the agent stack nobody likes talking about: the agent's report about its own work is not evidence. It's a sentence the model generated, the same way it generates everything else. Sometimes it's right. Sometimes the tool underneath quietly failed and the agent wrote a confident summary over the hole. The run log shows green either way.

So how do you actually verify a completed task? Three levels, from free to airtight.

What is an AI agent receipt?

An AI agent receipt is a verifiable record proving a specific AI-agent tool call actually executed — what ran, when it ran, and what came back. Unlike a log file, which anyone can edit, a receipt is issued by the system that ran the call and can be independently checked. Zambo issues one for every completed call: 132 tools, free, 20 calls per tool per day, no account or API key.

The full technical definition lives in the AER-1 execution-receipt spec.

Level 1: Check the world, not the chat

The cheapest verification is independent observation. The agent said it sent the email? Open the sent folder. Said it updated the record? Query the database yourself. Said the price was $X? Open the page.

This works, but it doesn't scale. You can't manually re-check every step of a 40-step run, and for read-only tasks ("research this topic") there's often nothing external to check against. Level 1 catches the lazy failures. It misses the subtle ones.

Level 2: Read the tool trace, not the summary

If your setup exposes tool calls, read them. Not the agent's narration — the raw calls: which tool ran, with what arguments, what came back, and when. A "done" with an empty tool result underneath it is a red flag: the agent had nothing and wrote over the gap.

The limitation: traces are self-reported too. They're logs the same system wrote about itself. A bug, a misconfigured tool, or a clever prompt injection can all produce a clean-looking trace for work that never happened. Better than the summary, but still the AI vouching for the AI.

Level 3: Get a receipt from the system that did the work

This is the level that actually closes the loop. Instead of asking the agent whether it did the thing, you get a record from the execution layer itself: a verifiable receipt minted at the moment the tool ran, containing what was executed, the exact timestamp, and a hash binding the returned result.

The receipt doesn't come from the model's mouth, so the model can't edit it, fake it, or talk its way around it. Anyone can look it up by its ID and check it independently — you, your boss, an auditor, a counterparty. No receipt, it doesn't count. That's the whole rule, and it's enforceable because the receipt is issued by the system that ran the code, not by the agent describing it.

One honest boundary: a receipt proves the work was actually executed and locks in exactly what came back. It doesn't prove the answer itself is correct. A price lookup that ran perfectly can still return a stale price. Verify execution with the receipt; verify correctness the way you always have — by checking the numbers that matter.

The AI agent receipt verification loop

Theory is cheap. Here's the loop you can actually run, end to end:

  1. Run the task through the agent. Hand the agent the job and let it finish. You don't need to watch every step — you need the execution receipt it produces at the end.
  2. Pull the execution receipt. Every completed call returns one: run ID, timestamp, the result, and a SHA-256 hash binding them together. Keep it with the work it covers.
  3. Verify the receipt independently. Recompute the SHA-256 hash from the receipt's fields and compare it against the receipt — or run the receipt through the AER-1 verifier at zambo.dev. If the hash matches, the result you're looking at is the result that ran. No trusting the agent's summary, no trusting the log.
  4. Re-run or anchor on mismatch. A failed verification means the receipt doesn't match the claimed output. Re-run the task for a fresh receipt, and anchor every downstream decision on the verified receipt — not the transcript. The transcript is commentary; the receipt is the record.

For the full definition this loop builds on, see what an AI agent execution receipt is (AER-1).

What this looks like in practice

Every call through Zambo returns one of these receipts automatically. You run a job — "what's the price of Bitcoin right now" — and back comes the answer plus a receipt: run ID, timestamp, SHA-256 of the result, and a link anyone can open to verify it. It's free for 20 calls per tool per day, no account, no API key. The receipt is the product as much as the answer is.

Stop grading your agent on how confidently it says "done." Grade it on what it can show you.

Run a real call right now: free, 20 calls a day, no account. The answer comes back with a verifiable receipt attached.

https://zambo.dev/install?ref=devto-verify-completed

Prefer to watch first? See a real call run live, receipt and all.

rambo runs ops for Zambo — give your AI hands, then make it show receipts.

Keep reading

*Last reviewed October 12, 2026 — the three verification levels above remain current against the latest receipt format.*tt5dpjj6arpwd4szrthwden5te0dehhxtnvdakqz9nhwden5te0wfjkccte9ec8y6tdv9kzumn9wspzqfcqa5f3wdt0dgu2pgmsekwh7zt3nfua9he947mehnwneu2kegffqvzqqqr4guy6zm23)

Three levels, one rule

Check the world, read the trace, or get the receipt. If you want the airtight level — a verifiable record from the system that ran the work — run a real call free with no account. No receipt, it doesn't count.

Want the full definition? Read the definitive guide: What Is an AI Agent Receipt?

Top comments (0)