Large swarms of AI agents may finish work faster than one agent, but Toby Ord’s analysis suggests the speed comes with diminishing returns and substantially higher total compute. The useful takeaway is less “more agents equals smarter” and more “parallel work has a coordination tax”.
Watch Desk analysis
What happened
In Swarm Scaling, Toby Ord examines how large groups of AI agents scale against traditional inference-scaling approaches such as making one model spend longer reasoning. The analysis uses simulated data derived from OpenAI’s GPT 5.6 launch results, rather than a report of a new real-world swarm deployment.
Ord models the problem with a “stepping-on-toes” measure for task efficiency and parallelisability. The idea is familiar from human teams: adding people can increase output, but overlapping work and coordination eventually eat into the gains. AI swarms appear to face a similar curve.
Why it matters
Swarms can offer a major practical advantage in elapsed time because many tasks run simultaneously. That could matter for research, software work and other workloads where a queue of independent subtasks is more painful than the compute bill.
The catch is that matching a single agent’s performance at scale may require significantly more total compute. Faster answers are not automatically cheaper answers, and a swarm that spends its time duplicating effort is merely a very efficient way to create an expensive pile of near-misses.
Our read
Ord’s analysis is a useful argument for measuring AI systems on both quality and resource use. “How quickly did it finish?” is only half the question. Teams also need to ask how much compute was consumed, how much work overlapped and whether the tasks were genuinely parallelisable.
The evidence is simulated, so this is a scaling model and a research prompt, not a universal operating rule. Still, it offers a sharper lens than the usual agent enthusiasm: sometimes the winning architecture is a swarm, sometimes it is one capable worker with enough time to think, and sometimes it is a spreadsheet quietly preventing both from setting fire to the budget.
What to watch
- Whether real agent systems reproduce the modelled diminishing returns.
- Which software and research tasks benefit most from parallel execution.
- How production teams report quality, latency and total compute together.
- Whether coordination methods reduce duplicated or conflicting agent work.
Discussion spark: Should AI teams optimise first for the fastest answer or for the lowest total compute, and who should decide when the two goals conflict?
Sources and evidence
- Swarm Scaling – Toby Ord (21 September 2026, 19:51 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.