agent-org

An agent team for your repo.
You merge, until it earns more.

Agents find bugs, fix them and open PRs. Reviewers can't edit code, a guard hook blocks the shortcuts, and every miss a human catches becomes a test.

See a run View on GitHub

install, in Claude Code

/plugin marketplace add BbekShr/agent-org
/plugin install agent-org@agent-org

Then, inside your project:

/org-init
/<project>-ceo audit

About 4 seconds to install. /org-init reads the repo and works on a branch; audit is a read-only report. Stop anything with touch .claude/ceo/PAUSE.

4production agent orgs it was distilled from
28principles, each with its reason, many traced to a real incident
23agent roles in the catalog, staffed from your repo's real seams
10/10first eval case: a real bug a human had found by hand. One case, not a benchmark

One run, one team, start to finish.

You invoke the CEO. Read-only specialists find, one editor fixes, an adversarial board and the release gate judge, and a PR comes back to you. The team shown is what /org-init generates for a standard Next.js + Supabase app with a few bots, with the model each role runs on.

ONE RUN, WITH THE TEAM /org-init GENERATES FOR A NEXT.JS + SUPABASE APP The CEO is a skill in your session: it orchestrates and never edits code. Model per role: judgment on the top tier, building and reviewing on the mid tier, mechanical work on the small tier. You /<project>-ceo Preflight kill switch, level, tree Assess specialists, parallel Decide re-verify, rank, cap Build one editor, one branch Review + gate frozen tree PR evidence section Scouts + checkers read-only, in parallel security-engineer sonnet silent-failure-hunter sonnet design-critic opus repro sonnet findings at file:line Plan for hard items built-in, read-only Plan agent opus the plan Builders the only roles that write frontend-engineer sonnet backend-engineer sonnet data-engineer sonnet the diff Review board + gate read-only, adversarial reviewer (per lens) sonnet on hard diffs opus release-manager haiku CONFIRMED defects Who merges by autonomy level L0-L2 you L3+ agents, you sample Records: memory, scorecard, lessons, evals every run writes them before the PR; the next run's preflight reads them PR outcome: landed, edited, reverted next run a revert or an escaped defect demotes the level Guard hook on every tool call blocks: gh pr merge --admin · force push · --no-verify · edits to gate and org control files · during a run: merge below L3, push to base, git stash
Roles come from the repo's real seams, models from the job: judgment (design critique, the plan for a hard item) on the top tier, building and reviewing on the mid tier, the mechanical gate on the small tier. Defaults top = opus, mid = sonnet, small = haiku; /org-init asks before fixing them. A small library gets three agents; a large repo with public issues adds researcher, architect, bug-hunter and rnd.
  1. Only builders write code.The CEO, the scouts, the reviewers and the release manager have no edit tools, so nobody grades their own work.
  2. Green gates are necessary, not sufficient.Every diff faces one reviewer per lens and the release gate together, on a frozen tree.
  3. Trust is earned.Promotion needs scorecard evidence and your say-so. A revert demotes the org on its own. Only you raise the level.

Full role catalog · The run cycle · The guard

One CEO, ten modes.

/<project>-ceo <mode>. Every mode starts with preflight and ends with a report; apply modes end at a PR. The autonomy level decides which are open.

auditread-only · L0The report-only specialists write a board report. Nothing changes. Start here.
lightPR · L1 · defaultFast assess, then fix the top 1 to 3 items.
deepPR · L1Every specialist, every finding surfaced, fixes batched by owner.
<ID>PR · L1One named backlog item through the full pipeline.
request "<ask>"PR · L1Plain words from anyone, engineer or not, into a spec, then built.
triage "<bug>"records or PR · L0Reproduce the report on main first. Then confirm fixed, file it, or fix it.
discoverrecords · L0Hunt and file new bugs, one area and one lens at a time. discover fix also fixes them (L3).
rndrecords · L0Propose features and process changes with evidence. Never builds in the same run.
issuesrecords · L0Triage open GitHub issues, through the untrusted-input screen, into backlog items.
e2ePR · L1Drive the running app: find, fix, re-drive, ship. Local only, never production.

The last four are optional; /org-init asks which to include. Stop any run with touch .claude/ceo/PAUSE.

Autonomy is earned, one level at a time.

An org starts where its tests can prove a change correct. Promotion needs scorecard evidence and your approval. A revert demotes it on its own.

L0

Observe

Audits, bug discovery, proposals. Records only.

L1

Verify

Fix PRs, one at a time. You merge.

L2

Constrain

Adds scheduled headless runs.

L3

Parallelize

Parallel worktrees. Agents merge green PRs; you sample after.

L4

Cloud

Always-on routines, same gates.

Most orgs start at L1. Only you raise the level. Every level keeps the same gates and the same review board: a level unlocks capabilities, never shortcuts.

Measured, not claimed.

Agents are tested like code. /org-eval checks out the commit before a real fix and asks the agent to review it, then a different model scores the answer.

This is the first run on photobari: a swipe bug in its photo viewer that every existing agent missed and a person found by hand the next morning.

10/10

Reviewer, as generated

Found the bug with the right lines and the fix that later shipped, plus three more defects that fix addressed.

5/10

Same reviewer, without the org's lessons

Right mechanism, wrong trigger. The lesson is what makes the difference, and the score shows it.

$ python3 .claude/ceo/bin/org_doctor.py .
info   autonomy L1 (target L2), 8 agents, 7 gates
info   feedback: off
warn   learnings/_shared.md: ## Active has 48 lines (>40)

0 error(s), 1 warning(s)

Start with an audit.

Generate the org on a branch, run /<project>-ceo audit, and read the board report. It changes nothing.

Get started