Vol. X · Field Operations · X.5
Verification as Architecture
The product of a coding agent is not code. It is code that passed a verifier the harness ran. Tests, typecheck, build, browser smoke, and a human gate are stages of the loop, not a ceremony after it.
7 min read
Doctrine
- Every act loop has a verify tool: the test command from AGENTS.md. Failure is an observation, not a shrug.
- Write the verifier before the feature when you can. Spec-driven is this idea with a document.
- Golden evals (20–50 real tasks) exist for the agent product itself, not only for the model.
Verifier in the loop
OpenHands and Codex run commands and read the output back. promptfoo and Inspect are the batch cousins. Evaluator–optimizer is the same shape for non-code artifacts. The architecture is: propose, verify, compact the error, propose again, cap the budget. If verify is missing, you built a typewriter.
Anti-patterns
- An agent that never runs the test command.
- Eval-by-demo on a laptop.
- Treating linter noise as success.