AgentOps Assurance Profile

AgentOps provides an evidence-producing workflow for AI-agent-created work. It is an engineering aid, not a certification, approval, release authority, data boundary, or guarantee of correctness.

Proven floor

• behavior-first intent with explicit scope and evidence requirements;
• one bounded implementation experiment;
• deterministic content identity through subject-manifest.v1;
• one independent judgment from a distinct fresh context;
• a durable PASS | FAIL | NOT_PROVEN verdict with checked and unchecked scope.

The freshness attestation is a declared trust fact, not cryptographic proof of model isolation. A consumer may require stronger runtime controls, multiple judges, or formal review for high-consequence work.

Consumer responsibility

The consumer owns model/provider approval, data handling, retention, secrets, Git and CI policy, tracker state, human approvals, release, rollback, and any formal assurance mapping. AgentOps packets and verdicts may be inputs to those systems; they do not replace them.

Data handling

Treat plans, candidates, evidence, and verdicts as potentially sensitive. Choose their storage, retention, redaction, and export policy before using them in a controlled environment. The default verdict directory is local to the workspace and can be overridden by the caller.