Subagents & Orchestration
The agent can spawn specialized helpers and batch tool calls with code mode.
For larger jobs the agent can split work across subagents and batch many tool calls efficiently with orchestration.
Subagents (roles)
The agent can spawn focused helper agents, each with a role that defines its mode, allowed tools, and iteration budget. Built-in roles include:
| Role | Mode | Good for |
|---|---|---|
| explorer | browser | Quick scouting across pages |
| researcher | browser | Deeper, multi-step web research |
| worker | agent | Carrying out a concrete task |
| reader | agent | Reading and extracting from content |
The main agent spawns them, waits for results (individually or all together), checks status, and can stop them. Roles are Markdown files with frontmatter, so you can add your own.
Example: "research the top 3 vector databases and summarize trade-offs" — the agent may spawn a researcher per database and combine their findings.
Orchestration (code mode)
Instead of issuing tool calls one at a time, the agent can use orchestrate to
write a short Python script that calls many tools in one pass. The script runs in a
sandboxed interpreter (no imports, no file system, no network beyond the provided
tools), with automatic checking and repair. This is much faster for repetitive,
multi-step work (e.g. visiting 20 pages and collecting one field from each).
If you want to review before it runs, the agent will produce a plan first.