swarm-advanced

Initialize a swarm with swarm_init specifying a topology and max agents, then spawn specialized agents with agent_spawn and orchestrate work via task_orchestrate. Alternatively, run npx Codex-flow swarm with strategy and mode flags from the CLI.

Use mesh topology for research and analysis since agents communicate peer-to-peer. Use hierarchical topology for development workflows needing a coordinator, star topology for centralized testing, and ring topology for sequential pipeline processing.

Yes, parallel_execute runs independent tasks concurrently across spawned agents. Use sequential orchestration only when tasks have dependencies, and monitor execution with swarm_monitor at a set interval.

Use memory_persist with a session ID, organize data with memory namespaces and TTL values, and create state snapshots with state_snapshot. You can restore later via context_restore and back up memory stores to a file path.

Coordination failures usually stem from wrong topology selection, missing memory configuration, or disabled monitoring. Verify the topology matches your workflow, check memory namespaces, and enable swarm_monitor to observe agent health.

Parallel execution fails when tasks have hidden dependencies, resource limits are exceeded, or error handling is absent. Identify truly independent tasks first, implement fault tolerance with auto-recovery, and analyze bottlenecks before scaling agent count.