Example: Multi-Agent Supervisor

import { Aside } from ‘@astrojs/starlight/components’;

The feedback loop pattern — the defining feature of OrgLoop. Claude Code sessions emit actor.stopped events, which route to a supervisor actor that reviews work and can re-dispatch tasks. The supervisor’s own completions feed back into the system, creating a recursive loop.

What this example shows

actor.stopped events as first-class routing signals
The recursive loop: actor completes -> event emitted -> routed to supervisor -> supervisor completes -> event emitted -> …
SOPs (Standard Operating Procedures) as launch prompts that guide actor behavior
Why actor.stopped is deliberately neutral — the system observes, actors judge

The pattern

Claude Code session ends
       |
       v
  actor.stopped event
       |
       v
  Route: session-review
       |
       v
  Supervisor agent (reads payload, evaluates work)
       |
       +-- Work complete? Move on.
       +-- Work partial? Re-dispatch with instructions.
       +-- Session crashed? Investigate and retry.
       |
       v
  Supervisor session ends -> actor.stopped -> ... (loop continues)

The loop sustains itself. Every actor completion is an event. Every event can be routed. The supervisor is just another actor whose completions feed back through the same system.

Prerequisites

Node.js >= 22
OrgLoop CLI installed (npm install -g @orgloop/cli)
A GitHub account with a repository to monitor
Claude Code installed locally (for the post-exit hook)
An OpenClaw instance running locally — OpenClaw is an AI agent orchestration platform that manages agent sessions

Setup

1. Environment variables

Variable	Description	Where to get it
`GITHUB_REPO`	Repository being worked on (`owner/repo`)	Your GitHub repo URL
`GITHUB_TOKEN`	GitHub PAT with `repo` read access	GitHub Settings > Tokens
`OPENCLAW_WEBHOOK_TOKEN`	Bearer token for OpenClaw API	Your OpenClaw instance
`OPENCLAW_DEFAULT_TO`	Default message recipient	Your OpenClaw team configuration

2. Install the Claude Code hook

orgloop hook claude-code-stop

This registers a post-exit hook so Claude Code emits actor.stopped events when sessions end.

3. Run

cd examples/multi-agent-supervisor
orgloop validate
orgloop start

The config

`orgloop.yaml`

# orgloop.yaml — Multi-Agent Supervisor
# Claude Code sessions feed back through a supervisor actor, creating a recursive loop.

apiVersion: orgloop/v1alpha1
kind: Project
metadata:
  name: multi-agent-supervisor
  description: "Claude Code supervision with feedback loop"

defaults:
  poll_interval: 5m
  log_level: info

# ─── Sources ─────────────────────────────────────────────────────────────────

sources:
  - id: claude-code
    description: Claude Code session completion events
    connector: "@orgloop/connector-claude-code"
    config:
      hook_type: post-exit
    emits:
      - actor.stopped

  - id: github
    description: GitHub PR activity (work produced by Claude Code)
    connector: "@orgloop/connector-github"
    config:
      repo: "${GITHUB_REPO}"
      token: "${GITHUB_TOKEN}"
      events:
        - "pull_request.review_submitted"
        - "pull_request_review_comment"
    poll:
      interval: 5m
    emits:
      - resource.changed

# ─── Actor ───────────────────────────────────────────────────────────────────

actors:
  - id: supervisor
    description: Supervisor agent — reviews work, re-dispatches if needed
    connector: "@orgloop/connector-openclaw"
    config:
      base_url: "http://127.0.0.1:18789"
      auth_token_env: "${OPENCLAW_WEBHOOK_TOKEN}"
      agent_id: supervisor
      default_channel: slack
      default_to: "${OPENCLAW_DEFAULT_TO}"

# ─── Routes ──────────────────────────────────────────────────────────────────

routes:
  - name: session-review
    description: "Claude Code session ended -> Supervisor reviews the work"
    when:
      source: claude-code
      events:
        - actor.stopped
    then:
      actor: supervisor
      config:
        session_key: "orgloop:claude-code:session-review"
        wake_mode: now
    with:
      prompt_file: "./sops/review-session.md"

  - name: pr-review
    description: "PR review received -> Supervisor handles feedback"
    when:
      source: github
      events:
        - resource.changed
      filter:
        provenance.platform_event: pull_request.review_submitted
    then:
      actor: supervisor
      config:
        session_key: "orgloop:github:pr-review"
        wake_mode: now
        deliver: true
    with:
      prompt_file: "./sops/review-pr.md"

# ─── Loggers ─────────────────────────────────────────────────────────────────

loggers:
  - name: console-log
    type: "@orgloop/logger-console"
    config:
      level: info
      color: true

  - name: file-log
    type: "@orgloop/logger-file"
    config:
      path: ~/.orgloop/logs/supervisor.log
      format: jsonl
      rotation:
        max_size: 50MB
        max_age: 7d

Routes

Route	Trigger	Actor	SOP
`session-review`	Claude Code `actor.stopped`	supervisor	`sops/review-session.md`
`pr-review`	GitHub `resource.changed` (PR review submitted)	supervisor	`sops/review-pr.md`

The SOP: `sops/review-session.md`

SOPs are launch prompts delivered to the actor alongside the event. They tell the actor what to do with the signal it received.

# Review Claude Code Session

A Claude Code session has completed. Review what happened and decide next steps.

## Instructions

1. Read the session payload — check `exit_status`, `duration_seconds`, and `summary`
2. If a summary is provided, evaluate whether the described work sounds complete
3. Check the working directory for any uncommitted changes or failing tests
4. **Work looks complete** -> verify the PR is ready for review, move the ticket forward
5. **Work is partial** -> identify what remains and re-dispatch with specific instructions
6. **Session crashed (non-zero exit)** -> investigate the failure, retry with a simpler approach
7. **Session was very short (<30s)** -> likely hit an error early; check logs and retry with fixes

## Guidelines

- Don't trust the summary blindly — verify against actual repo state
- If re-dispatching, be specific about what needs to happen next
- Escalate to a human if the same task has failed multiple times
- Keep a running count of retries; stop after 3 attempts on the same task

The supervisor reads the actor.stopped payload (exit status, duration, summary) and cross-references it with actual repo state. It does not trust the completing actor’s self-report.

Why `actor.stopped` is neutral

OrgLoop does not interpret session outcomes. It observes that a session ended and routes that fact. The actor.stopped event payload carries details — exit code, duration, summary — but OrgLoop does not assign meaning to them.

This is deliberate. Whether work was completed, the agent crashed, got stuck, or fabricated its summary — that is for the receiving actor (the supervisor) to judge. OrgLoop routes signals. Actors have opinions.

How the loop sustains itself

Claude Code finishes a coding session. The post-exit hook fires.
The Claude Code connector emits an actor.stopped event.
The session-review route matches and delivers the event to the supervisor.
The supervisor reviews the work using the SOP as guidance.
The supervisor’s own session eventually ends — emitting another actor.stopped event.
If a route matches supervisor completions, the loop continues. If not, the chain terminates naturally.

The loop is not infinite by design. It terminates when no route matches the latest actor.stopped event, or when the SOP instructs the supervisor to stop retrying.

When to use this pattern

Code review automation: Claude Code writes code, supervisor validates it
Multi-step workflows: break complex tasks into sessions, with a supervisor coordinating handoffs
Quality gates: every AI-produced artifact passes through a review step before merging
Retry with escalation: failed sessions get re-dispatched with more context, up to a retry limit