Multi-Agent Orchestration

For complex jobs that fan out across many parallel sub-tasks, Chief can spawn worker agents and coordinate them.

When to Use It

Single-agent execution is the right default. Spawn workers when:

• The work is embarrassingly parallel (e.g., process 1,000 invoices, one per worker).
• The work has independent sub-tasks that can run concurrently and be merged.
• You want time-bounded completion regardless of total work volume.

A typical multi-agent job:

Chief, here's a list of 200 customer accounts. For each one,
pull their last 90 days of usage, score risk, and write a
one-paragraph summary. Run them in parallel and give me a
single report when all are done.

Chief shows a live status in the thread:

Spawning 200 workers across 4 batches. Estimated time: 3-5 minutes. Tracking in this thread.

▰▰▰▰▰▰▱▱▱▱ 124 / 200 done

Configuring Concurrency

By default, Chief uses up to 10 parallel workers per task on Enterprise. You can raise or lower this:

Chief, run that with 25 workers in parallel.

Or set a tenant-wide cap in Settings → Multi-Agent → Max concurrency.

Cost & Tokens

Multi-agent jobs multiply token usage. Each worker is a separate context. Chief estimates the total cost up front and shows it in the thread before kicking off:

200 workers × est. 4,200 tokens each = 840,000 tokens ($2.50 at current rate). Reply with go to start.

Failure Handling

Worker failures are isolated — one bad input doesn’t kill the whole job. Chief retries each failed worker up to 3 times, then includes the failures in the final report with a clear “could not complete” note.

Visibility

Dashboard → Tasks → [task] shows the full worker DAG: each sub-task, its tokens, its tool calls, its result. Useful for auditing or debugging a job that produced unexpected output.

Limits

• Max concurrent workers per task: 50 (raise via support ticket if you need more)
• Max workers per tenant per minute: 200
• Max worker depth (workers spawning workers): 2 levels