AI teams: How a lead Agent coordinates specialized SubAgents
Split complex work into bounded specialties instead of making multiple Agents repeat the same exploration.
Parallelism is not the default answer
Parallel work saves time only when a task can be divided into boundaries that do not compete for the same file, state, or decision. Once one Agent has resolved a path, another should not repeat it from the beginning.
The lead Agent owns context and acceptance; specialized Agents own bounded outputs
- Declare the files, modules, or evidence owned by each workstream first.
- Share one completion condition so each workstream does not invent a different definition of “done.”
- After results return to the lead workstream, deduplicate them, check conflicts, and then enter the release gate.
- No Agent may resolve a conflict by overwriting someone else’s changes.
Distinguish available capabilities from Preview
Specialized SubAgent collaboration is available now. Persistent AI team templates and personality-based role recommendations remain Preview and must not be described as formally delivered.
This is also a basic rule for Kyormar content: plans, demonstrations, and production capabilities are described separately.