SwarmCoordinator
Agent Orchestration
The SwarmCoordinator manages agent lifecycle, work distribution, and health monitoring. Each agent runs through the model-agnostic AgentHarness, supporting Claude Code, Codex, Gemini, browser agents, or custom models. Default claim strategy is best-fit.
Autonomous Task Worker (AgentHarness)
_
SwarmCoordinator subscribes to Agent Mail, starts heartbeat every 30s, begins work loop.
Get ready tasks, filter by skills, sort by priority, claim highest match.
Retrieve memory context, spawn AgentHarness with any LLM, capture output.
Store learnings in Memory System, publish completion via Agent Mail.
AgentHarness Lifecycle
Model-Agnostic Execution
AgentHarness wraps any LLM CLI for task execution:
claude --print --dangerously-skip-permissions "<task prompt>"
The agent captures stdout/stderr, monitors for completion, and reports results back through Agent Mail.
Supported Skills
Agents can be configured with any combination of skills:
Claim Strategies
5
Claim strategies available
Any
LLM model supported
∞
Tasks per session
Configuration
Start agents via CLI with skill-based registration:
# Start swarm with 5 agents using smart prioritization jetpack start -a 5 --strategy graph-prioritized # Or join as a single agent with specific skills jetpack agent --skills typescript,react,frontend