Summary
Three agents, one task, one measured outcome. The lab that teaches whether multi-agent coordination is worth the coordination cost — and when a single agent is faster, cheaper, and more reliable.
Objectives
- 01Design a task that plausibly benefits from multiple agents.
- 02Implement three agents with defined roles and handoffs.
- 03Instrument the coordination — who calls whom, why, how often.
- 04Compare end-to-end against a single-agent baseline.
Key Ideas
- Multi-agent is a design choice, not a virtue; the baseline is the test.
- Every coordination message is a cost; count them.
Lab
Choose a task with defensible role separation (e.g., research → draft → critique; or plan → execute → verify). Implement three specialized agents with distinct tools and prompts. Ship a working orchestrator. Instrument every message between agents. Compare on a fixed test set against a single strong agent baseline. Report cost, latency, and quality. Write a decision memo: does multi-agent win here, or is the single-agent baseline strictly better?
Deliverables
- — Multi-agent system with three specialized agents.
- — Single-agent baseline for the same task.
- — Instrumentation of coordination messages and costs.
- — Comparison report with a defended recommendation.
Rubric
- — Role separation — the three agents do meaningfully different work.
- — Baseline — an honest single-agent comparison exists.
- — Instrumentation — you can answer 'where did coordination fail?'.
- — Judgment — the recommendation is defended, not asserted.