Part 3 of 4 — Multi-agent systems are a graph

The cheapest graph is one node.

Part 2 made fan-out look powerful. It is also expensive: more prompts, more traces, more places that need approval, and more connections that can pass on wrong information. OpenAI’s orchestration guide says it plainly: start with one agent whenever you can. Add specialists only when they materially improve capability isolation, policy isolation, prompt clarity, or trace legibility. Split only when the job really changes (different instructions, tools, or rules), not just because you want more agents.

That is the honest point here. A single loop is still a graph.

cheap:  [one agent] ──► reply

split:  [A] ──► [B]
          ↘──► [C] ──► merge
        only if the contract changes

Orchestration still decides when each step runs. You do not need a large team of agents just because others use one.

Agreement is not proof

When you do fan out, the merge must see the writes. It must not treat “both specialists agreed” as proof.

vote:   B yes + C yes  ──►  "done"

check:  B write + C write  ──►  merge
                                + test / schema / human

Why AI Agents Fail Part 4 already covers this: majority-style consensus inflates confidence, and models often err together. Disagreement is a signal. Agreement still needs an independent check — a test, a schema, a human who can refuse.

A human approval is itself a node with a stop: the edge does not fire until someone named clicks. Orchestration covers where that approval step sits. This part only says: do not skip it because two branches gave similar answers.

Prompts and loops still matter. Each node still needs a job description, and each branch still needs a stop. Part 1 drew the map. This part says it is fine to keep the map small.

Takeaway: Start with one agent. Split only when the instructions, tools, or rules change. A merge that only counts votes has not verified anything.

Sources