Multi-Agent Systems Are a Communication Problem

The intuition behind multi-agent design is organisational: give each agent a role, let them collaborate, and complex work decomposes the way it does in a good team. The intuition is right about decomposition and wrong about what makes teams work.

Roles are cheap, protocols are not

Naming an agent a Researcher and another a Critic costs a prompt. What actually determines whether the pair produces anything useful is the protocol between them: what is passed, in what shape, and what happens on disagreement.

Most disappointing multi-agent systems fail in one of three ways, and none of them is a reasoning failure:

• Context truncation between hops — agent three receives a summary of a summary and confidently works from a distortion.
• Unbounded consensus — two agents politely converge on the first plausible answer because neither is incentivised to hold a position.
• Silent role collapse — every agent gradually behaves like a generalist assistant, because the prompts differ but the training does not.

Pass structure, not prose

# fragile: the next agent re-reads and re-interprets English
{"result": "I found three issues, the main one is auth related..."}

# durable: the next agent consumes fields, not narrative
{"findings": [
   {"id": "F1", "severity": "high", "area": "auth",
    "claim": "session token not rotated on privilege change",
    "evidence": "src/session.ts:88", "confidence": 0.82}
 ],
 "coverage": ["auth", "session"], "unchecked": ["billing"]}

The second form survives a hand-off because there is nothing to misread. It also makes the failure visible: an empty `unchecked` array that should not be empty is a bug you can assert on.

Make disagreement structural

If your critic agent shares the generator's context window, it is not a critic. It is a second opinion from the same person.

Genuine adversarial review needs asymmetry — a different lens, a different slice of evidence, or an explicit instruction to refute rather than assess. Redundancy catches slips. Only diversity catches blind spots.