Yesterday-AI Agent Swarm Orchestration 🐝
Orchestrate specialized coding agents in isolated worktrees. Based on Elvis Sun's architecture.
Concept
- Orchestrator (You): Holds business context, decisions, memory. Spawns agents.
- Workers (Agents): Hold codebase context only. Run in isolated
git worktrees. - Loop: Monitoring script checks PRs, CI, and agent health.
Usage
1. Spawn an Agent
# Creates a worktree and prepares the task context
./skills/agent-swarm/scripts/spawn-agent.sh <task-name> <branch-name> "<prompt>"
2. Monitor Progress
# Checks all active agents and reports status
./skills/agent-swarm/scripts/monitor-agents.sh
Structure
Each agent gets:
- Isolated Git Worktree:
../worktrees/<feature-branch> - Task Context: A dedicated prompt file.
- Session: A tracked session (TMux or OpenClaw Sub-Agent).
Rules 🛡️
- CONTEXT_SEPARATION: Orchestrator knows why (Business). Worker knows how (Code). Do not overload workers with business docs.
- DEFINITION_OF_DONE: PR created + CI passed + Self-Review done.
- FAIL_FAST: If an agent is stuck, kill it and respawn with better context (don't loop).
Adapted for OpenClaw by ManniTheRaccoon. 🦝