hawkes.aiKey ideas
- The one idea — Adding agents has diminishing then negative returns: coordination overhead grows and errors compound across a chain, so past a point more agents make a system slower, pricier, and less reliable — not better.
- How it is shown — A task done well by three agents degrading as it scales to fifteen — hand-off delays pile up and one agent's small error propagates and corrupts the whole result.
- The trap to avoid — Believing "more agents = more intelligence" — beyond the right number, coordination cost and compounding errors dominate; often one strong agent beats a big swarm.
- What it sets up — The plumbing is powerful but has limits — and so does the whole thing. what still breaks?
Adding more AI agents can make your system worse. Coordination overhead grows, and every hand-off compounds errors — past a point, one strong agent often beats the swarm.
The one idea
Adding agents has diminishing then negative returns: coordination overhead grows and errors compound across a chain, so past a point more agents make a system slower, pricier, and less reliable — not better.
We've built up to swarms and subagents, and it's tempting to conclude the answer to any hard problem is just more agents. It isn't. There's a point where adding agents makes things worse — and it's worth understanding why. The first cost is coordination. Every agent you add has to be told what to do, and its results gathered and merged. That's communication, and it grows fast — more agents, more hand-offs, more places for information to get lost. At some point you coordinate more than you actually do. The second cost is sharper: errors compound. Each agent is a probabilistic model, with some chance of a small mistake. Chain ten, each feeding the next, and those errors multiply — one agent's wrong assumption becomes the next one's starting fact.
How it works — the demo
A task done well by three agents degrading as it scales to fifteen — hand-off delays pile up and one agent's small error propagates and corrupts the whole result.
A long chain can quietly corrupt its own result. So there's a curve. One agent to a few often helps a lot — real parallelism, real specialization. But keep adding, and the gains shrink, then reverse: coordination and compounding errors dominate. Past that peak, more agents buys overhead and fragility, not capability. It mirrors human teams. Two good people can outproduce one. But throw thirty at a small task and they trip over each other — meetings, miscommunication, mistakes multiplying. The answer is rarely the most people; it's the right number. Agents are no different.
The trap to avoid
Believing "more agents = more intelligence" — beyond the right number, coordination cost and compounding errors dominate; often one strong agent beats a big swarm.
Why it matters — and what’s next
The plumbing is powerful but has limits — and so does the whole thing. what still breaks?
The trap: believing more agents equals more intelligence. Past the right number, you're not adding brains — you're adding coordination cost and chances to fail. Very often, one strong, well-equipped agent beats a sprawling swarm. Reach for more only when the task genuinely needs them. So more agents isn't always better: coordination overhead and compounding errors cap the gains, and past a point make things worse. That's one honest limit of this plumbing — but not the only one. For all its power, what does the whole apparatus still get wrong? Next: the limits of memory and harnesses.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
We've built up to swarms and subagents, and it's tempting to conclude the answer to any hard problem is just more agents. It isn't. There's a point where adding agents makes things worse — and it's worth understanding why.
The first cost is coordination. Every agent you add has to be told what to do, and its results gathered and merged. That's communication, and it grows fast — more agents, more hand-offs, more places for information to get lost. At some point you coordinate more than you actually do.
The second cost is sharper: errors compound. Each agent is a probabilistic model, with some chance of a small mistake. Chain ten, each feeding the next, and those errors multiply — one agent's wrong assumption becomes the next one's starting fact. A long chain can quietly corrupt its own result.
So there's a curve. One agent to a few often helps a lot — real parallelism, real specialization. But keep adding, and the gains shrink, then reverse: coordination and compounding errors dominate. Past that peak, more agents buys overhead and fragility, not capability.
It mirrors human teams. Two good people can outproduce one. But throw thirty at a small task and they trip over each other — meetings, miscommunication, mistakes multiplying. The answer is rarely the most people; it's the right number. Agents are no different.
The trap: believing more agents equals more intelligence. Past the right number, you're not adding brains — you're adding coordination cost and chances to fail. Very often, one strong, well-equipped agent beats a sprawling swarm. Reach for more only when the task genuinely needs them.
So more agents isn't always better: coordination overhead and compounding errors cap the gains, and past a point make things worse. That's one honest limit of this plumbing — but not the only one. For all its power, what does the whole apparatus still get wrong? Next: the limits of memory and harnesses.