Definition
Plain language
Using a separate AI agent to actually run and test another agent's output instead of just reading it.
As stated in the literature
A verification approach where a distinct agent deploys and interacts with an artifact (e.g., a generated app via browser automation) to produce a reward signal, rather than judging from source or text alone.
Why it matters: It catches failures that look fine on paper but break when run, giving a far more trustworthy signal of whether an output truly does the job.
For example, instead of just reading the code of a generated web app, a separate agent opens it in a browser, clicks the buttons, and checks whether it actually works.
Heard on the show
“They call it agent-as-a-verifier.”Episode 090 — How MiniMax-M2 Bets That Sparsity Plus Verifiable Rewards Can Match Frontier Agents