Ten AIs are not smarter than one — they just lose more on the way
June 17, 2026 · Kevin Kelly, Out of Control~4 min read
Everyone "knows" the next leap in AI is teams — swarms of agents, each a specialist, conferring like a tiny company inside your laptop. More minds, more power. It's the most intuitive idea in the field right now, and a 2026 study just put a crack in it: in a fair fight, one good agent beat the whole committee. The reason is older than AI, and a man named Kevin Kelly wrote it down back in 1994.
The receipt on the swarm
This spring, Stanford researchers Tran and Kiela published a result with an unglamorous title and an awkward finding: under an equal "thinking-token" budget, a single LLM outperformed a multi-agent system on multi-hop reasoning. The culprit is mundane and can't be fixed by cleverness — every time one agent hands off to another, it has to compress what it knows into a natural-language summary, and every summary leaks. Researchers even have a name for the slow rot that follows: "agent drift," where a crowd of agents gradually wanders off-task over a long run. None of which has cooled the gold rush — multi-agent workflows on Databricks grew 327% from June to October 2025, and in February 2026 Anthropic shipped Agent Teams as a product. The market is buying swarms faster than the evidence can vet them.
Kevin Kelly saw the hive mind coming
Long before any of this, in his 1994 book Out of Control, Kevin Kelly went looking for the logic of systems made of many dumb parts — beehives, ant colonies, flocks, markets, the early internet. He called them "vivisystems," and he was half in love with them. A hive has no boss; intelligence emerges from thousands of simple agents following local rules, and the result is astonishingly robust, adaptive, and creative. This is exactly the dream the swarm crowd is chasing: don't engineer one genius, breed a colony and let intelligence emerge. Kelly's enthusiasm was real — and so was his warning, right there in the title.
What a swarm actually buys you
Because Kelly's central law is a trade, not a free lunch. The more you distribute a system, the more adaptive and resilient it gets — and the less you can predict or control it. You can't have the emergent magic without surrendering legibility; a hive is powerful precisely because no one is steering it, which is also exactly why no one can. A swarm of agents inherits the whole bargain: it can cover enormous ground in parallel and route around a single failure, but it will also drift, surprise you, and resist any clean audit of why it did what it did. The diagram below lays the two regimes side by side — the lone agent you can follow, and the swarm you can only herd.
A 2026 Stanford paper (Tran & Kiela) argued that under an equal thinking-token budget, a single LLM outperforms a multi-agent system on multi-hop reasoning — because every inter-agent handoff, passed as a natural-language summary, loses information; researchers also describe 'agent drift,' behavioral degradation over long multi-agent runs. Yet adoption is exploding: multi-agent workflows on Databricks grew 327% from June to October 2025, and Anthropic shipped Agent Teams with Claude Opus 4.6 in February 2026. Kevin Kelly's Out of Control names the trade: a swarm is a distributed, emergent system — resilient and broad, but never fully legible or controllable. Use many agents to cover ground in parallel, not as a reflex. A swarm buys breadth and pays in coherence.
The handoff tax
There's a children's game that explains the Stanford result better than any equation: telephone. Whisper a sentence down a line of ten kids and what comes out the far end is mangled — not because any child is dumb, but because every retelling drops a little. Inter-agent handoffs are telephone with a token budget. Each agent reads the previous one's summary, not its full reasoning, so detail evaporates at every hop; with enough hops, the swarm is confidently debating a problem that has quietly mutated since the first link. A single agent holds the whole thread in one head and never pays this tax. That's the unglamorous reason one mind so often beats the committee: it never has to translate itself.
When to actually call the swarm
So the lesson isn't "swarms are bad" — Kelly would object, and he'd be right. It's that a swarm is a specialized tool with a specific bill, not a default upgrade. Reach for many agents when the job is wide rather than deep — fan out across a hundred files, search in parallel, try a dozen approaches at once, survive a node dying. Reach for one strong agent when the job is a single chain of tight reasoning, where every handoff is a leak you can't afford. And whichever you pick, remember Kelly's bargain: every bit of autonomy you hand the swarm is a bit of control you don't get back. Don't summon a colony to do a thinker's job.
A swarm buys breadth and pays in coherence.
Every bit of autonomy you hand a colony is a bit of control you never get back.
Framework drawn from Kevin Kelly's Out of Control (凯文·凯利《失控》) — vivisystems, the hive mind, emergence, and the control-vs-adaptability trade. The single-vs-multi-agent finding is from a 2026 Stanford paper by Tran and Kiela (single agents outperforming multi-agent systems on multi-hop reasoning under an equal token budget; "agent drift"). Adoption figures (Databricks multi-agent workflows +327%, June–October 2025; Anthropic Agent Teams, February 2026) per popular tech reporting; figures per the original sources.
技术
十个 AI 不比一个聪明——只是在路上漏得更多
2026 年 6 月 17 日 · 凯文·凯利《失控》约 4 分钟
人人都「知道」AI 的下一跃是团队——一群智能体,各是专才,在你笔记本里像一家小公司一样开会。脑子越多,本事越大。这是眼下这个领域最符合直觉的想法,而 2026 年一项研究刚在它上面磕出一道裂缝——在一场公平较量里,单个智能体赢了整个委员会。原因比 AI 还老,一个叫凯文·凯利的人,早在 1994 年就把它写下了。