Split a task across parallel agents, each with its own model, tools, and step budget. Steer any of them while the swarm runs, and only their handoff summaries come back to the parent.
Every running agent exposes an inline interjection input. Send a message to the swarm orchestrator and it fans out to all active children within 3 seconds. Or target a specific child by its ID. The agent polls for operator messages every 2-3 steps via a dedicated check_operator_message tool.
Each sub-agent can run on a different provider. A researcher child on a fast cloud model, a coder child on your local Qwen model, a reviewer child on Claude for quality gates. Configured via SubagentProfiles with the standard provider/model format.
Up to 8 agents running concurrently with independent tool loops
Steer any agent mid-run without stopping the swarm
Different providers and models per sub-agent role
Completion ratio, quality pass rate, serial collapse detection
Blocked recursive delegation, configurable tool include/exclude per role
Only handoff summaries return to parent — no context pollution