Brain and worker architecture

Desktop splits large jobs between two roles: a brain that plans, changes files, and writes the response, and workers that take on broad reading. Because file contents don't enter the brain's context, token usage and time drop noticeably.

How it works

1. 1The brain plans the work and picks out the parts that need broad reading.
2. 2It hands those parts to workers with the delegate tool; up to 2 workers run in parallel per turn.
3. 3Workers run for up to 15 steps with read-only tools and return a short report.
4. 4The brain changes the files and writes the response based on the reports.

If a worker fails, work doesn't stop; the brain reads the files itself.

Models in Auto mode

Expert, Apex, and Auto are available only on the Plus and Max plans (see Credits).

Example scenario

When you say “Find every API endpoint in this repository and document them,” the brain plans the work and splits scanning the files between two workers with delegate. The workers scan the folder with read-only tools, list the endpoints, and return short reports. The brain writes the document from those reports; the full contents of the files never enter its context.

Frequently asked questions

• Do I need to pick the workers? No. The brain hands off broad reading itself with the delegate tool.
• Can workers change my files? No. Workers use only read-only tools; the brain changes the files.
• Does work stop if a worker fails? No. The brain carries on by reading the files itself.