Multi-agent workflows
Separate independent work while keeping responsibility for the final result clear.
Updated 2026-09-15
Use multiple agents when the assignments can proceed independently. Give one owner responsibility for combining the results. Additional agents are most useful when each has a concrete question or a bounded set of files.
Ordinary delegation creates tracked helper conversations. A managed Workflow is explicitly started and coordinates phases using its frozen role configuration. Asking for a second opinion does not require starting a managed Workflow.
Worked example: investigate before changing
Illustrative.
01 · Goal
Understand a bug's cause and the tests that cover it.
02 · Setup
Configure suitable roles and choose a workspace. Keep both helpers read-only so their work can run independently.
03 · Prompt
Use at most two helpers. One traces the reported failure through the code.
The other identifies existing tests and missing coverage. Both are read-only
and must not delegate. While they work, inspect the reproduction details.
Combine their findings into one proposed fix before editing.04 · Evidence produced
A caller trace, a coverage summary, and a proposed fix that reconciles both reports.
05 · User review
Check that the proposed fix addresses the reproduction. If implementation is then split, assign file ownership and remind each agent that the workspace is shared.
See Split discovery and review for a reusable assignment.