Multi-agent ACLs in one project

A pattern for fitting more than one worker (and a read-only critic) under a single manager — with channel membership doing the ACL work, not runtime gating. #product is the open team room; #internal is a back-channel for the two devs only; #all is the wildcard. The critic sits in permission_mode: plan and so cannot mutate anything it reviews.

The patterns here also show up organically in examples/newsletter-office/ (layered managers + workers) and examples/oss-maintainer/ (channel-isolated pipeline). The compose file below stays preserved as a reference recipe.

.team/projects/swarm.yaml

version: "2.0.0"

project:
  id: swarm
  name: Multi-Agent Swarm
  cwd: .

channels:
  - name: product
    members: [manager, dev1, dev2, critic]
  - name: internal
    members: [dev1, dev2]
  - name: all
    members: "*"

managers:
  manager:
    runtime: claude-code
    model: claude-opus-4-8
    can_dm: [dev1, dev2, critic]
    can_broadcast: [product, all]

workers:
  dev1:
    runtime: claude-code
    model: claude-sonnet-4-6
    reports_to: manager
    can_dm: [manager, dev2, critic]
    can_broadcast: [product, internal]
  dev2:
    runtime: claude-code
    model: claude-sonnet-4-6
    reports_to: manager
    can_dm: [manager, dev1, critic]
    can_broadcast: [product, internal]
  critic:
    runtime: claude-code
    model: claude-opus-4-8
    permission_mode: plan
    reports_to: manager
    can_dm: [manager, dev1, dev2]
    can_broadcast: [product]

The critic in permission_mode: plan is the same archetype as market-analysis’s risk dissenter — a read-only agent that proposes but never mutates.