Insights

Do Agent Swarms Really Work?

By Will Hardman · · 6 min read

Agent swarms are an exciting and trendy way to attack difficult research or analysis problems. The big idea is that having several, independent agents collaborate makes it more likely you’ll see a good solution - much like having a diverse group of people collaborate on a task can lead to better outcomes.

OpenAI’s recent breakthrough on the Navier-Stokes problem is the poster-child for a multi-agent system (MAS). The exact architecture of the system is, to my knowledge, unknown outside of the company. They describe 10,000 “coordinating agents” divided into groups with each group focused on a specific variant of the problem. During the run, they “cross-pollinated” findings between the groups. Anthropic’s work on the Erdős problems earlier in 2026 apparently used a similar “summarise and cross-pollinate” technique.

Outside of the labs, researchers and practitioners have experimented with a variety of MAS set-ups. There are as many taxonomies as there are articles on MAS but for the purposes of this article, I’ll group them into three:

  1. The orchestrator-worker pattern. Here, a central “orchestrator” agent splits a problem up and tasks independent “workers” with solving the parts. The orchestrator then collates and integrates the results.
  2. The routing pattern. In this pattern, a set of specialised agents are defined. A task is handed to one of them. The agent processes the task as best it can (according to its specialisation) and then, if it remains incomplete, hands it to another agent for further work.
  3. Teams and swarms. This pattern subsumes a multitude of different organisations with a common theme: a number of parallel agents - with shared resources - tackle a problem concurrently. Usually they exchange information via a central “blackboard” or via a “project manager” agent.

The thing that all three have in common is non-deterministic routing - by which I mean we do not know in advance exactly how many tasks each agent will perform and in what sequence.

From a practitioner’s perspective, this makes MAS difficult to debug, observe, evaluate or optimise. Furthermore, there is a well-established “communication overhead”: MAS are token-hungry beasts and it simply costs more to use them. In other words, we give up control and accept additional costs in exchange for flexibility and - we hope - capability.

If we possibly can, it is always preferable to design our systems as single agents or as agentic workflows - rather than as MAS. An agentic workflow is a system where we specify, in advance, a graph of agents and operations through which a task flows. In other words, we constrain the system’s flexibility and retain a lot more control. This is exactly what the langgraph framework is designed to support and why I am such a big fan of it.

What the evidence says

You might wonder, therefore, when it makes sense to ditch the strictures of a graph. After all, a quick scroll through your LinkedIn timeline will soon present you with an exciting claim about an agent swarm with some kind of super-intelligent capabilites.

A review of the recent literature gives us a pretty clear answer to the question, which I will summarise below in three lessons.

Lesson 1. Gains are mostly explained by tokens.

Often, most of the gains from MAS are explained by the simple fact that they consume more tokens working on a problem than single agents do. Where MAS architectures are compared with a single agent under a token-controlled baseline, the advantages from the multi-agent set-up disappear - or even reverse - except in very specific circumstances (see Lesson 2, below). Reports and case studies that do not report against this baseline should be taken with a large pinch of salt.

Lesson 2. The orchestrator-worker pattern is good for parallelisable, wide, read-only tasks

The orchestrator-worker pattern has clear benefits when the task can be characterised as “parallelisable” (it can be broken into sub-tasks to be run concurrently), “wide” (the main challenge is covering a large number of simpler tasks) and “read-only” (no sub-agent needs to write to a shared state). Research tasks can be good examples of this. The reason this pattern works on these problems is now well-understood:

  1. Each sub-agent has a clean, small context relevant to its task which makes it easier for the sub-agent to attend to and reason over the information.
  2. Running the sub-agents in parallel reduces the wall-clock time for the task to complete. If a sub-agent stalls or returns a sub-standard result, the orchestrator can re-run it without impacting the rest of the task.

So, the orchestrator-worker pattern can offer advantages when the task is complex enough to “fill up” a sizable fraction of a single agent’s context window.

Lesson 3. Don’t bother with routing, swarms or teams.

There is limited-to-no evidence for the superiority of routing architectures and swarms / teams. I have yet to see any published research that compares these architectures to a single agent in a token-controlled manner and demonstrates an improvement. Without this evidence, no agent-builder should pick these patterns, given the known issues with cost and control that will ensue.

In conclusion

So what are we to make of the stunning results in mathematics reported by the labs with their large-scale, multi-agent setups?

Firstly, note how mathematical problems can be broken down into smaller problems and tackled independently - they have parallelisable and wide properties so 10,000 agents al benefit from the smaller context windows and the entire system runs with a faster wall-clock time.

Secondly, since sub-agents’ results can be formally verified with proof engines, there is a very clean way to collect successful sub-results. (Note this is not true for open, non-verifiable research tasks.) This crucial property of mathematical proof enables the designers to collate and push verified information to the sub-agents mid-search. What is disclosed feels very much like an orchestrator-worker pattern.

Thirdly, note that neither Anthropic nor OpenAI published a token-matched, single-agent comparison for their results. (Understandable, but the absence of such a baseline makes it difficult to tell what contribution the architecture made to the outcome.)

In my view, these results are really telling us something about the power of inference-time scaling, rather than about agentic architectures per-se. Simply spending more tokens on a problem - and using a MAS pattern to keep the context-windows lean and informative, lets a system conduct a massive, parallel search and, eventually, to stumble upon the answer.

For the rest of us, building agentic systems to work in a messy world, it’s best to scroll past the hype-posts about swarms and concentrate on designing good, agentic workflows where we can. It’s the best-evidenced route to success.

Ready to move from pilot to production?

Book a free 30-minute consultation. You'll leave with a straightforward answer on where agentic AI can work for your organisation, and the first steps to get there.

Book a free 30-min consultation →