Self-Verification (AI)
Self-verification is a technique where an agent checks its own output against criteria — tests, schemas, rubrics — and retries if the check fails.
In plain English
Self-verification is the agent version of "proofread your work." After producing an output, the agent runs a check: does the code compile? does the JSON validate? does the answer cite the source I claimed? If the check fails, the agent uses the failure as feedback and tries again. The check can be deterministic (run a test, run a linter, match a schema) or model-based (use an evaluator prompt).
The technique is powerful because it converts a single noisy output into a verified one, at the cost of a few extra turns. The best agent workflows combine cheap deterministic checks (runs in milliseconds) with occasional model-based evaluators (runs in seconds). That layering catches most issues without ballooning latency.
Why it matters for Black Box
Black Box's Coding specialist self-verifies by running tests and linters after every patch. The Content specialist self-verifies brand voice and fact-citation against a rubric. The Business Ops specialist self-verifies numeric outputs against sum checks. Self-verification is why the product can ship changes without a human review on every step.
Examples
- A coder running
pnpm testafter a patch and retrying on failure. - A writer checking that claimed stats match the cited source.
- A planner validating its output against a JSON schema before returning it.