In a short post, Toby Ord proposes treating AI "swarms"—multiple agents running in parallel—as a new form of inference-scaling. The trade-off he highlights is that swarms often require more total tokens than a single agent, but their parallelism can finish tasks faster in wall-clock time.
Key findings
-
Ord identifies speed as the primary advantage of swarms. A four-agent swarm needed roughly twice the total number of tokens to reach the same performance as a single agent, but each agent used about half as many tokens. Because the agents run in parallel, the same task can theoretically be completed in roughly half the wall-clock time.
-
Scaling a swarm is not perfectly linear. As the number of agents grows, returns diminish in a way analogous to coordination costs observed in human teams. Economists call this the "stepping on toes" parameter: adding more participants raises coordination overhead.
-
Concretely, increasing agents by a factor of ten does not yield tenfold performance; instead the gain is 10^λ times, which Ord reports translates in practice to roughly 3–5×. This shortfall accumulates with larger scale-ups, so a large swarm can fall progressively behind an equivalent token budget used by a single agent.
-
Despite the diminishing returns, swarm scaling may be potent enough to raise the probability of a rapid, RSI-like (recursive self-improvement) intelligence explosion rather than lower it. Ord had hoped the λ parameter for AI agents would be smaller—making an intelligence explosion less likely—but his observations do not support that hope.
Why this matters
So far, AI capabilities have mostly been scaled by choosing the right mix of compute and data during training and then allocating inference budget to longer chains of thought and tool calls. Agents introduce an additional degree of freedom: how to divide inference work across multiple parallel, interacting entities.
If we learn to make agents coordinate productively—especially if they can achieve greater-than-sum-of-parts outcomes (as in notable coordinated incidents tied to platforms like HuggingFace)—the returns to swarm scaling could grow. Monitoring how swarms scale and coordinate will therefore be important for understanding the overall trajectory of AI capability growth.
Conclusions
Ord's analysis highlights that agent-based parallelism adds a new axis to AI scaling: it buys speed at the cost of total token efficiency. The diminishing returns due to the "stepping on toes" effect limit those gains, but do not necessarily prevent swarms from increasing the risk of rapid capability gains. Tracking and studying swarm coordination and efficiency will be crucial for anticipating future AI risks and opportunities.
Source
- Toby Ord: "Swarm Scaling" (post)
