boshu2/agentops

The operating loop a coding agent follows — and skills to orchestrate multi-agent systems.

What it solves

AgentOps addresses the unreliability of autonomous coding agents by solving four key issues:

  1. Self-grading bias: Agents often declare victory in the same session they wrote the code.
  2. Shared blind spots: A single context may share the same errors as the author.
  3. Requirement drift: Without fixed acceptance criteria, an agent's definition of "done" can change mid-task.
  4. Lack of traceability: Standard chat histories are ephemeral and difficult to audit.

How it works

The project implements a structured operating loop known as RPI (Plan $\rightarrow$ Implement $\rightarrow$ Validate) that runs as skills inside existing coding agents like Claude Code or Cursor. It uses "Beads" to lock in intent via BDD and DDD before implementation begins. To ensure integrity, the validate phase runs in a fresh context—potentially using a different model—to issue a verdict of PASS, FAIL, or NOT_PROVEN. It also includes admission-control hooks to prevent agents from executing destructive commands.

Who it’s for

  • Developers using AI coding assistants (Claude Code, Codex, Cursor) who want more reliable, verifiable code generation.
  • AI Engineers building multi-agent systems who need orchestration skills like swarm or council for independent judging.
  • Teams requiring auditable, content-addressed evidence of agentic work through persistent verdicts.

Highlights

  • Separation of concerns: Decouples the agent that writes code from the agent that validates it.
  • Multi-agent orchestration: Includes skills for managing fleets of agents and complex workflows.
  • Deterministic guards: Features a policy dispatcher to block destructive commands.
  • Verifiable evidence: Writes content-addressed, plain JSON verdicts to track agent performance.
  • Plug-and-play: Can be installed as a plugin for popular AI coding tools or via npx.