AgentOps

AgentOps helps fallible coding agents produce work that another fresh context can independently judge.

The product boundary is deliberately small:

RPI -> Plan -> Implement -> fresh Validate -> durable verdict -> report and stop

Proven floor

AgentOps provides:

• behavior-first intent with normal and edge acceptance examples;
• one bounded RED -> GREEN -> refactor experiment;
• deterministic content identity independent of Git;
• one fresh, author-distinct semantic judgment;
• a standalone, content-addressed PASS | FAIL | NOT_PROVEN verdict.

No fresh evidence-backed verdict means the experiment is not proven. The verdict's context separation is explicitly attested; it is not claimed as cryptographic proof of model isolation.

What AgentOps does not own

AgentOps is not a new GitLab, CI service, tracker, merge queue, delivery system, release manager, scheduler, or autonomous retry controller. It does not own:

• retries, budgets, queues, claims, leases, work ownership, or next actions;
• Git commits, branches, pushes, PRs, merges, rollback, closure, or release;
• the caller's decision after a verdict;
• mandatory provenance or learning on the validation critical path.

Repositories and callers keep those policies. They may use direct pushes, PRs, hosted CI, merge queues, cloud agents, or custom release systems without asking AgentOps for delivery permission.

Product surfaces

Four load-bearing skills define the core:

Skill	Responsibility
rpi	dispatch Plan, Implement, and Validate once; report and stop
plan	shape acceptance, evidence, and write scope
implement	run one bounded experiment and produce a candidate
validate	independently judge exact content and persist the verdict

learn remains an optional off-path consumer of verdict collections. Strategy skills such as Premortem, Postmortem, Council, and idea genies add judgment when the caller wants it. Factory/runtime adapters such as NTM, Agent Mail, Gas City, and swarms may provide roles and dispatch. None is a correctness or lifecycle authority.

The ao CLI supplies deterministic repository checks and generic read-only or record helpers where useful. Semantic validation belongs to the Validate skill, not a CLI state machine.

Sovereign evidence

The durable verdict is plain JSON under caller-controlled storage. It binds acceptance, exact subject content, author and validator identities, criterion results, evidence, checked scope, and omissions. A generic provenance ledger may copy or reference it later, but ledger availability is never required for validity.

Learning thesis

The long-term hypothesis remains that recurring validated mistakes can become better context, tests, or deterministic checks. That compounding claim is not the core completion boundary. Learning runs off-path, cites distinct verdicts and findings, and proposes evidence for later evaluation; it does not silently promote policy or control another experiment.

Why this shape

Coding agents are stochastic and can overstate completion. More orchestration does not itself create trust. The smallest useful trust boundary is an explicit behavior, exact content identity, and a fresh independent judgment whose limits are recorded. AgentOps packages that boundary without taking over the user's engineering system.