← All drillsOrchestrate a Multi-Agent System
What you'll be able to do- ✓Distinguish supervisor-led orchestration from peer-initiated delegation
- ✓Model a multi-step task as a dependency graph instead of an ordered list
- ✓Add a third worker that depends on two upstream results and wire its inputs correctly
- ✓Build a supervisor that sequences workers over A2A while holding no business logic itself
- ✓Identify and exploit independent steps in a plan to run them in parallel, with a measured latency gain
- ✓Handle a worker's mid-plan failure with a reasoned retry-or-abort decision instead of a crash
- ✓Produce a structured, human-auditable trace of a full multi-agent run
Coordinate several independently-built agents into one system that completes a real multi-step task end to end — the exact shape a larger autonomous-engineering system needs at full scale.
⌁ Multi-agent frameworks with role-specialized agents under a shared plan — MetaGPT and AutoGen are the best-known published instances of this coordination pattern.
Start this internshipCreate an account to unlock the 9 sections, the workbench, and AskThili.
BeginSections
1. Orchestrate a Multi-Agent System
🔒 locked2. Lesson 1 — From delegation to supervision
🔒 locked3. Lesson 2 — The plan as a dependency graph
🔒 locked4. Lesson 3 — The third worker
🔒 locked5. Lesson 4 — Wiring the supervisor
🔒 locked6. Lesson 5 — Exploiting independence
🔒 locked7. Lesson 6 — A worker fails mid-plan
🔒 locked8. Lesson 7 — The decision report
🔒 locked9. Lesson 8 — Batch measurement
🔒 lockedDig deeper
📄MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework (Hong et al., 2023, arXiv:2308.00352)
paper📄AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (Wu et al., 2023, arXiv:2308.08155)
paper🔗thili-ai/thiliruntime — the reference implementation you build in this course
code