AgentOps
Autonomous code validation for coding agents
Coding agents can produce plausible code that is still wrong. AgentOps helps answer the two questions that decide whether you can trust the work: is the code right, and is the agent output proven enough to grant more autonomy? It sits on top of the agent you already use (Claude Code, Codex, Cursor, OpenCode) and adds the validation membrane, evidence trail, and repo-local corpus that make that judgment repeatable.
See it work
/discovery → bead graph · /crank → sub-agents in waves · /validate --mixed → real Claude + Codex verdict. Live sessions. MP4
AgentOps breaks intent into bounded slices, gives each a failing test and a write scope, and makes every phase boundary a gate that records evidence. The agent starts loaded with prior decisions and learnings instead of cold:
> /council --mixed validate this PR
[council] evidence sealed → 6 judges across Claude Code + Codex CLI
[claude/judge-1] WARN rate limiting missing on /login
[codex/judge-1] WARN token bucket lacks jitter under burst
[claude/judge-2] PASS redis integration follows pattern
Consensus: WARN, fix /login limit + refill jitter before shipping
Recorded → .agents/council/<run-id>/verdict.md
What you get
The center is validation: prove the agent output, keep the proof, and use that record to decide how much autonomy the next run earns. The supporting layers all stay local in .agents/ (no telemetry, no hosted control plane):
| Layer | The problem | What AgentOps adds |
|---|
| Validation membrane | agent output can look correct while being wrong | tests, local gates, /pre-mortem, /vibe, /council, and pawl verdicts prove or reject the work |
| Evidence trail | "looks good" does not survive handoff | .agents/ captures runs, decisions, findings, citations, verdicts, retros, and closeout proof |
| Context compiler | validators and implementers start cold | ao context assemble builds phase-scoped packets; ao lookup retrieves decay-ranked knowledge |
| Knowledge ratchet | lessons vanish between sessions | /forge mines learnings, /evolve reconciles, and durable lessons become constraints before more autonomy is granted |
The corpus is an LLM wiki of markdown. Agents read it natively and write to it as they work, so it maintains itself instead of becoming another doc you keep up by hand. Public citations of measurable flywheel or corpus outcomes use promoted artifacts under docs/evidence/ (e.g. 2026-04-02 flywheel case study); .agents/ remains the local operating substrate. Why that beats Notion or Confluence: docs/wiki-for-agents.md. The full theory (context as the lifecycle, the CDLC): docs/cdlc.md.
Install
Pick your runtime, then type /quickstart in the agent.
# Claude Code
claude plugin marketplace add boshu2/agentops
claude plugin install agentops@agentops-marketplace
# Codex CLI (macOS/Linux/WSL). OpenCode: install-opencode.sh
curl -fsSL https://raw.githubusercontent.com/boshu2/agentops/main/scripts/install-codex.sh | bash
# Codex CLI (Windows):
irm https://raw.githubusercontent.com/boshu2/agentops/main/scripts/install-codex.ps1 | iex
# Gemini / Antigravity
curl -fsSL https://raw.githubusercontent.com/boshu2/agentops/main/scripts/install-agy.sh | bash
# Other skills-compatible agents
npx skills@latest add boshu2/agentops --cursor -g
The ao CLI is optional but recommended (bookkeeping, retrieval, health, the loops):
brew tap boshu2/agentops https://github.com/boshu2/homebrew-agentops && brew install agentops # macOS
# Windows: irm https://raw.githubusercontent.com/boshu2/agentops/main/scripts/install-ao.ps1 | iex
# Or release binaries / build from source (cli/README.md).
Installs hookless: skills and the ao CLI guide the workflow, and the local cockpit gate is the release authority. GitHub Actions are an optional/manual backstop, not the routine shipping path. The only hard requirement is an agent runtime and git; everything else degrades gracefully. Full dependency matrix: docs/dependencies.md. Day-2 install, update, backup, permission, recovery, and escalation paths are in docs/install-day2-ops.md.
Quick start
