swarm-coordination

Multi-agent coordination decomposes tasks across workers using swarm patterns. Initialize with swarmmail, decompose the task into subtasks, spawn workers via swarm_spawn_subtask, track progress with swarm_status, and aggregate results after all agents complete. Swarm Mail is mandatory for all messaging and reservations.

Workers must follow nine mandatory steps: initialize, query learnings, load skills, reserve files, do work, report progress, checkpoint, store learnings, and complete. Each step uses specific tools like swarmmail for communication, semantic_memory_find for discovery, and swarm_progress for tracking to ensure coordinated execution and result collection.

Use swarm_decompose to break tasks into parallel work units targeting specific subsystems. Validate decomposition with swarm_validate_decomposition, then spawn subtasks with swarm_spawn_subtask. Track cross-subsystem progress via swarm_status and swarm_progress, using swarmmail for inter-worker communication and file reservations.

Yes. Swarm coordination supports knowledge gathering via semantic_memory_find, cass_search, and pdf-brain_search, then generates plans with swarm_plan_prompt. Workers execute in parallel, report findings through swarmmail checkpoints, and aggregate results before completing, enabling complex multi-step reasoning across agents.

Workers checkpoint progress regularly and store learnings to enable recovery. If a worker fails, swarm_status detects the issue, and swarm_review validates partial results. Checkpoints preserve state so remaining workers continue or retry—swarmmail reservations prevent conflicting file access during recovery.

No. Swarm Mail handles mandatory file reservations and releases across workers, preventing conflicts. Each worker reserves files before work, reports progress through swarmmail, and releases after checkpointing, ensuring safe parallel execution without manual locking.