Can an LLM verify another AI agent's work?
Yes, if it sees the right inputs. In our tests an LLM verifier that saw only the executing agent's report caught none of the tasks done on the wrong target. The same verifier, given the user's original request as well, caught 28 of 32 and cut false acceptance from about 20% to 2%. What the verifier is shown matters more than which model it is.
Answered from GDF Research experiments by Sterling Morrison. Numbers link to the papers below.
Why the agent's own report isn't enough
In most multi-agent setups, the executor finishes a task and returns a summary, and the summary is what the reviewer gets. If the executor deployed to shop.harborline.co instead of shop.harborline.com, every piece of evidence it gathered is about the .co site, and it all checks out.
We ran this as a fault-injection study: 48 scenarios in eight domains, each correct completion mutated five ways. Verifiers handed only the executor's report passed every wrong-target completion whose correct twin they also passed (0 of 11 caught for Claude Haiku 5.5, 0 of 6 for Claude Sonnet 5.5). The stronger model did slightly worse. They also rejected 25–54% of correct work over embellishments in the summary, such as "customer has been notified," that the evidence didn't cover.
What fixed it
Adding the user's request, about 30 words, moved false acceptance from 18–22% to 2% and removed every false rejection. Extra evaluation rules on top added almost nothing. Letting the verifier write its own acceptance criteria made it worse in the other direction: it rejected 75–88% of correct work, because it invented requirements the user never stated.
| Verifier receives | Haiku 5.5 | Sonnet 5.5 |
|---|---|---|
| Executor's report only | 18% | 22% |
| + the user's original request | 2% | 2% |
The other condition: it has to look
When verifiers had to choose whether to open the evidence log, those that opened it accepted 3% of faulty work and those that skipped it accepted 59%. Whether a verifier looks depends on how it's scored. Paid to agree with a panel that had already approved, it opened the log 6–9% of the time.
