Clawd Agent OS — AI skill for Claude Code

clawd-agent-os

A production-grade Python framework for orchestrating 50+ Claude agents in tiered hierarchies. Battle-tested across a real autonomous AI organization running 24/7 on a single Mac.

What this is

A blueprint for **multi-agent operating systems** at scale — not a toy, not a demo. This is the exact architecture used to coordinate a 54-agent system across health, finance, recruiting, research, and more, with:

• Tiered authority (T0 human → T1 CEO orchestrator → T2 domain orchestrators → T3 specialists → T4 workers)
• Durable inter-agent messaging that never silently loses messages
• Git-worktree-isolated fleet spawning so parallel agents can't stomp each other's files
• Model-tier routing with a single registry as the source of truth
• Location-agnostic queue draining — any machine, any number of workers, no global ceiling

Architecture

┌─────────────────────────────────────────────────────┐
│                    T0  Human                        │
│                    T1  CEO Orchestrator (Opus)       │
│          T2  Domain Orchestrators (Sonnet)           │
│       T3  Specialists (Sonnet/Haiku)                 │
│    T4  Workers (Haiku — ephemeral task runners)      │
└─────────────────────────────────────────────────────┘

Each tier can only be addressed through proper channels. Agents delegate down, verify up. No tier skips another.

Key components

Fleet orchestrator (`clawd_agent_os.fleet`)

Spawns Claude agents as isolated tmux sessions on git worktrees. Each run gets:

• A UUID4 run_id
• An isolated git worktree on task/ so file edits can't collide
• A WORKPLAN.md the agent re-reads each loop (Manus pattern — durable state)
• A headless claude --session-id -p --output-format stream-json
• A 6-hour wall-clock ceiling before auto-reap

from clawd_agent_os.fleet import Fleet

fleet = Fleet(repo_root="/path/to/repo")
handle = fleet.spawn(
    agent="researcher",
    object