Building multi-agent systems that don't collapse

Why most multi-agent systems fail

The same way most distributed systems fail: unbounded fan-out, no clear contract between actors, and a coordinator that quietly becomes a single point of failure.

A working contract

Every agent in the system needs three things written down:

1. A goal it owns, expressed as a single sentence.
2. An interface, the tools it can call and the schema of what it returns.
3. A budget, tokens, time, and recursion depth.

Without all three, you don’t have agents. You have very expensive while-loops.

The coordinator pattern

A thin coordinator that only routes and never does the work itself tends to outlive every “smart” orchestrator we’ve tried.

async function coordinate(task: Task) {
  const plan = await planner.run(task);
  for (const step of plan.steps) {
    await dispatch(step); // workers do the work
  }
}

Things we keep relearning

• Determinism at the edges, fuzziness in the middle.
• Logs are the product. If you can’t replay a session, you can’t debug it.
• Humans should be a tool the agent can call, not a fallback hidden behind an alert.

Closing

Multi-agent systems don’t earn their keep by being clever. They earn it by being legible.