Multi-Agent Systems, and When Not to Build One

Splitting a hard problem across specialised agents is genuinely powerful and usually premature. Daniel Okafor on coordination cost, the failure modes nobody warns you about, and the far simpler thing to try first.

In this episode

The microservices rhyme

Every argument for and against multi-agent systems was made twenty years ago about services. The lessons transferred; the industry did not read them.

Coordination is the whole cost

Handoffs, shared state, and the compounding error rate that turns four reliable agents into one unreliable system.

When it genuinely wins

Real parallelism, genuinely separable expertise, and independent verification. Three conditions, all required.

Try this first

One agent, better tools, tighter evaluation. It solves more cases than the conference talks suggest.

Two agents that disagree is a system. Two agents that disagree with no arbiter is an outage.

Mentioned

• The full written companion piece in the articles archive
• Module material on this topic in the AI Fluency programme

---

Beyond the Prompt publishes a new episode most weeks. Subscribe wherever you listen, or read the written edition first — both cover the same ground, at different speeds.