Capabilities

Agent swarm

Agent-swarm mode lets a lead model apply one prompt template to many items, run the child agents in parallel, and receive one ordered result. It is built on Roder's existing subagent dispatcher, so children inherit the parent workspace handle, policy boundary, cancellation signal, and authorized tool surface.

Use it from the TUI

/agent-swarm on
/agent-swarm status
/agent-swarm Review each changed file and report one risk
/swarm off

Persistent mode adds the swarm reminder to each turn without changing the text in your transcript. A one-shot command asks the model to use agent_swarm for that prompt. Mode is scoped per thread, so enabling it in one conversation does not change sibling threads sharing the same runtime.

Tool contract

{
  "description": "Review the changed files",
  "prompt_template": "Review {{item}} and report one concrete risk.",
  "items": ["src/api.rs", "src/runtime.rs", "src/tui.rs"]
}

The prompt template must contain the exact {{item}} placeholder.
Rendered prompts must be distinct and the full swarm is validated before any child starts.
Results stay in input order even when children finish out of order.
Failed or aborted children include resumable agent ids when a child reached running state.
agent_swarm must be the only tool call in a model response, preventing mixed partial work.

Bounded scheduling and rate limits

The scheduler starts a small initial burst, then paces later launches. You can cap active children, set per-child timeouts, and limit the swarm to at most 128 agents. Provider rate limits trigger exponential retries for the affected child and a shared capacity governor for the whole swarm, preventing every child from retrying at once.

[agent_swarm]
max_subagents = 64
initial_launch_limit = 5
launch_interval_ms = 700
max_concurrency = 4
child_timeout_seconds = 180
rate_limit_max_retries = 4
rate_limit_base_backoff_ms = 3000

Cancellation preserves completed output, marks started children as aborted, and records children that never launched separately. After a quiet period without more rate limits, global capacity recovers gradually instead of jumping back to full parallelism.

Live progress for every client

The TUI renders live completed, failed, and aborted counts plus a compact child result grid. App-server and SDK clients can toggle mode withthread/set_agent_swarm_mode and listen foragentSwarm/started, agentSwarm/progress, andagentSwarm/completed notifications. Per-child detail continues through the existing subagent-trace event family.

Policy stays explicit. Swarm mode does not bypass approvals or sandbox rules. When the current TUI policy would gate parallel child work, Roder asks for the policy switch before entering the mode.

For durable teammate threads, canonical task paths, follow-up work, and non-destructive interruption, see agent control.