In a world full of agentic factories, it's better to work on the factory, rather than in it.
One source, every agent
AgentRig treats AGENTS.md + .agents/rules/ as the single
source of truth and compiles it into Copilot, Claude, Cursor, Codex, OpenCode, and MCP.
Edit once; every surface updates.
Safe on existing repos
init is non-destructive by default — your existing AGENTS.md,
.mcp.json, and rules are preserved verbatim.
Details →
What you get
A turnkey harness built around 12 principles from production agent systems — and a compiler that projects it into every surface.
AGENTS.md as source of truth
Plain markdown. Glob-scoped reflex rules in .agents/rules/. No DSL.
Compiles to every surface
Projects into Copilot, Claude Code, Cursor, Codex, OpenCode, and VS Code MCP in one command. Surface map →
12 principles, scaffolded
State machine, role prompts (triager / developer / reviewer / judge / security-reviewer), skills, rules, wiki — all editable. Read them →
Evals you can run
Install-completeness + quality probes (deterministic), plus fixture-based agentic eval
with an independent judge and paired sign-test lift. eval --scaffold even
generates the scenarios from your repo's stack — answer "is this harness paying
for the tokens it spends?" with a real verdict. How →
Live dashboard
Terminal or HTML. Agent roster, live GitHub tasks per harness label, audit score — offline.
No lock-in
Local files, MIT licensed, no hosted service. Switching primary agents is a config change, not a rewrite.
Pick a starting point
init, see what lands in your repo. Five minutes.
The 12 principles →
The opinionated playbook AgentRig encodes.
Commands reference →
init, compile, update, doctor,
eval, dashboard.
Agent surfaces →
Which files project where, and the symlink layout.
Evaluating the harness →
3 layers — install completeness, quality probes, fixture-based agentic eval with
sign-test lift. Honest about what each does and does not prove.
Source on GitHub →
doidor/agentrig — issues, discussions, editable knowledge.