Run agent swarm
run_agent_swarmEngineer a system with a swarm of agents (ADR 0027): an orchestrator divides the description into subsystems and frozen interface contracts, divides again any subsystem too large for one agent (max_depth levels, default 2), specialists write the subsystems in parallel, and every answer is checked in the context of the whole model, with conflicts sent back to the agent that caused them. With a prompt alone, the result is a new saved model holding the skeleton and one federation proposal per subsystem. With model_id (a saved m_ model), nothing is divided: one agent develops each subsystem the model already has, keeping what is there, and its work is staged as a proposal on that same model; pass subsystem to go one level deeper, decomposing that subsystem's parts, and repeat level by level. The prompt is then optional guidance. Nothing is written until you review and commit each proposal. Paid. Call it directly: it charges what the job uses, never more than its estimated maximum or the max_spend_credits you set. request_id is optional; set one to make a retry safe. Returns job_id immediately. Poll get_job_status for later progress.
Tool call template
This tool declares no required parameters. Add optional parameters from the input schema as needed.
{
"id": 1,
"jsonrpc": "2.0",
"method": "tools/call",
"params": {
"arguments": {},
"name": "run_agent_swarm"
}
}Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| rows | No | Only for cost_mode "agent_authored" on generate_safety_analysis: {conditions:[{target,description,severity,...}],modes:[{target,mode,severity,occurrence,detection,...}]}; with analysis "stpa": {losses:[{id,description}],hazards:[{id,description,losses}],controllers:[{target,processModel}],controlActions:[{from,to,name,ucas:[{id,type,context,contextVariables,hazards,constraint}|{type,notApplicable,rationale}]}],feedback:[{from,to,name,updates}],constraints:[{id,name,text}]}. | |
| model | No | ||
| study | No | ||
| events | No | ||
| format | No | ||
| prompt | No | ||
| stop_s | No | ||
| analysis | No | For generate_safety_analysis: fha, fmea, both (default) or stpa. stpa needs no standard. | |
| fidelity | No | ||
| model_id | No | A model from a generation job (model_ id) or a saved studio model (m_ id); a saved model is read at the head of main. | |
| provider | No | ||
| standard | No | ||
| cost_mode | No | managed | |
| max_depth | No | Only for run_agent_swarm: levels of division, 1 to 3 (default 2). At 2 and 3, a subsystem the orchestrator judges too large for one agent is divided again inside its frozen boundary. | |
| subsystem | No | For reprompt_agent: qualified name of the subsystem, e.g. "Rover::DriveUnit". For run_agent_swarm on a saved model: the subsystem to decompose one level further (its parts get the agents); leave it out to work on the whole system's subsystems. | |
| request_id | No | ||
| scenario_id | No | ||
| simulations | No | ||
| specification | No | Only for cost_mode "agent_authored" on generate_system_model and build_full_system; see those tools for its schema. | |
| simulation_type | No | ||
| max_spend_credits | No |