Exercises — Module 04: Control flow and subagents

4.1 — Subagent vs. tool

Take an existing tool that produces a long free-form output (e.g., a web scraper or a file-read tool over a large file). Convert it into a subagent that returns a compressed summary.

Measure:

• Parent context size at turn N under each approach.
• Total cost.
• Task-success rate on a 10-case eval.

When is the subagent worth it?

Deliverable: results + recommendation.

4.2 — Router + specialists

Design a router for a help-desk bot with three specialists: technical, billing, account. Implement it. Route 50 real (or realistic) tickets. Measure routing accuracy.

Deliverable: confusion matrix + one paragraph on the hardest misclassifications.

4.3 — Critic loop

Take a writing task (e.g., “write a PR description for this diff”). Add a critic subagent that reviews the worker’s output and requests revisions. Cap revisions at 3.

Run on 20 examples. Is quality higher? At what cost?

Deliverable: numbers.

4.4 — Budget negotiation

Implement an escrow budget system: parent allocates N turns to child, child returns unused portion. Demonstrate that a runaway child cannot starve the parent.

Deliverable: code + a deliberately-runaway test case.

4.5 — Graph review

Take a multi-agent system from the open-source world (or design one on paper). Draw its agent graph. Identify one simplification and one elaboration that would make sense. Defend both.

Deliverable: diagram + analysis.