Module 25 · 45 min

Subagents

You can justify each subagent you add with one of the four documented benefits, cap it, and predict what comes back to the parent when it finishes.

Workbench tag
module-25
Draft

This page is not written yet. The outline below is the plan of record from curriculum/BUILD_TRACK_PLAN.md; the prose, the code, and the verified block land with the workbench’s module-25 tag.

What this page will cover:

  • When a subagent earns its place, the subagents page lists four benefits: context isolation, parallelization, specialized instructions, and tool restrictions. That last one is the ability to hand a reviewer Read and Grep and nothing that writes.
  • Programmatic AgentDefinition as the recommended creation path: only the final message returns to the parent, and maxTurns is the cap, because no per-subagent wall-clock deadline exists.
  • The hosting page’s warning about wide parallel fanouts and rate limits, batch instead of blasting everything at once.
  • CLI agent teams and SDK subagents are different features. The progression runs from one researcher to implementer plus reviewer to parallel research.
  • Ship: an optional review-fanout mode in the workbench, with the module 24 cost readout so you can see the tradeoff.