Some problems are too hard for a single model in a single pass — not because the model is weak, but because the problem needs sustained, structured effort. The emerging answer is multi-agent systems: decompose the problem, delegate pieces to sub-agents, coordinate them over time, and verify the result. It's how AI is starting to crack problems that resisted single-shot approaches.
The pattern
A coordinating process breaks a big problem into sub-problems, assigns each to an agent (or a fresh model call with focused context), and combines their results — often iterating, critiquing, and retrying. It mirrors how a team of humans tackles something hard: divide the work, let specialists focus, integrate, and check. No single agent holds the whole problem; the system does.
A single model has one shot and a fixed context. A multi-agent system has as many shots, as much focused context, and as much verification as you're willing to pay for.
Why it beats a single call
Three reasons. Focus — each sub-agent works on a bounded piece with relevant context, avoiding the "lost in the middle" and overload problems of cramming everything into one prompt. Parallel exploration — different agents can try different approaches. Verification — agents can check each other's work, and machine-checkable outputs (like formal proofs) let the system keep only what verifies. Errors get caught instead of compounding.
The cost
This is expensive — many model calls, lots of inference, long runtimes. It's test-time compute scaled up: trade money and time for capability. So it fits high-value, hard problems (research, complex engineering), not cheap high-volume tasks. The economics only work because capable models are getting cheaper.
Why it matters
Multi-agent decomposition is one of the most important ideas in current AI, because it extends what's possible, not just what's convenient. It's how systems tackle decade-old research problems and multi-hour projects. Understanding the pattern — decompose, delegate, verify — helps you see where AI capability is genuinely expanding, and how to structure your own hard problems for it.
