Multi-Agent Systems · ~10 min
Eight lessons, one reflex: default to one agent, and add coordination only where the task structure forces it. Here's the whole course as a lookup table — and a mixed review to prove it stuck.
The through-line of the whole course: coordination is a cost you pay to buy independence or verification — never a free upgrade. Reach for the smallest structure the task actually demands.
| Symptom | Move | From |
|---|---|---|
| A single agent struggles on a hard task | Climb the ladder one rung; reach multi-agent only when simpler levels stop being reliable — budget ~15× tokens | L1 |
| Work splits into independent directions | Orchestrator-worker — decompose, parallel workers with scoped tools, then synthesize (not concatenate) | L2 |
| High-stakes design needs the best of several attempts | Fan-out N=3–5 with engineered diversity, then a synthesis agent + committee review | L3 |
| A child must disagree with the parent | Spawn fresh — fork only to extend a load-bearing decision with siblings to amortize | L4 |
| Information is lost between pipeline stages | Typed handoff — conclusions not transcripts; gate the first write with plan approval | L5 |
| An agent declares "done" on broken work | Verify-gated completion — independent read-only verifier on the critical path, fail closed | L6 |
| The system misbehaves and you want a bigger model | Diagnose the MAST bucket first — 41.8% of failures are design; rewrite prescriptive prompts as framework prompts | L7 |
| You need to ship a prompt/model change in production | Rainbow deploy — N versions drain independently; roll back the whole tuple via a router change | L8 |
Part 1 was when and how: the complexity ladder gates the whole decision, and the two canonical shapes — orchestrator-worker (one task into subtasks) and fan-out synthesis (one task, N attempts) — both live or die on the synthesis step, never a concatenation. Part 2 was the contracts that hold a fleet together: fork-vs-fresh on each spawn, a typed handoff plus plan gate between stages, and admission control over every "done." Part 3 was operating the fleet: the failure taxonomy that tells you to suspect the design before the model, and rainbow deployment for shipping changes without breaking live sessions. One move, applied at each layer — pay for coordination only where the structure earns it.
Every lesson had a backfire box, and they rhyme: most multi-agent configurations lose to a strong single agent. Fan-out on a routine task, a plan gate on a one-file edit, a verifier with no ground truth, a rainbow deploy for single-turn Q&A — each adds cost with no return. Reach for coordination when the task genuinely decomposes or a mistake is expensive, not on reflex.
Mixed review — across all eight lessons
Question 1 · from L1The default reflex before reaching for multiple agents is to…
Question 2 · from L2An orchestrator that concatenates worker outputs adds…
Question 3 · from L4For a child that must reset bias and disagree, you spawn…
Question 4 · from L6In admission control, an ambiguous "done" claim resolves to…
Question 5 · from L8A deployable agent version should be rolled back as a tuple of…