boshu2/agentops

What it solves

AgentOps solves the problem of "self-grading" in agentic engineering, where the same AI agent that writes code also declares it finished and correct. It prevents rubber-stamping and acceptance drift by separating the authorship of code from its independent judgment, ensuring that a fresh, independent context validates the work against the đúng original intent.

How it works

AgentOps acts as an operations layer that integrates with existing coding agents (like Claude Code, Codex, or Cursor). It introduces a deterministic judgment step where a fresh reviewer context—often from the same model family but in a new session—evaluates a candidate change. This reviewer issues a binding verdict of PASS, FAIL, or NOT_PROVEN based on an evidence contract that requires a subject manifest, coverage of changed paths, and distinct author/validator IDs.

It provides an optional library of skills (e.g., plan, test, refactor, validate) that agents can use to perform specific tasks. It also includes a deterministic CLI (ao) for checks and evidence operations, and admission-control hooks to block destructive commands.

Who it’s for

It is designed for developers and engineers building agentic software factories or using AI coding agents who require rigorous, verifiable, and durable evidence of code quality and correctness.

Highlights

• Separation of Concerns: Decouples the agent that implements a change from the agent that judges it.
• Evidence Contracts: Uses content-addressed verdict.v2 files in external storage to create a durable, replayable record of judgment.
• Skill Library: A comprehensive menu of modular skills for research, planning, implementation, and validation.
• Factory Agnostic: Supports various software factory runtimes, including Gas City and the Agentic Coding Flywheel.
• Deterministic Guardrails: Includes policy dispatchers to prevent agents from executing known-destructive commands.