Photo: Tara Winstead / Pexels
Researchers report around 70% accuracy on SWE-bench by combining narrow agents built by different developers
The AI industry has spent years assuming that a bigger, smarter single model is the answer. A new paper from Princeton University suggests a well-organized committee may beat the lone genius.
The researchers found that combining specialized agents, each built independently by different developers, improves overall performance beyond what a single general-purpose agent can manage. Their multi-agent setup reached around 70% accuracy on demanding benchmarks covering software engineering and reasoning tasks.
That matters now because the race to build better AI agents is crowded. OpenAI, Anthropic, and Google DeepMind are all competing hard on agent performance. A finding that favors coordination over sheer scale could shift how that race gets run.
How the marketplace of agents works #
The core idea is something the authors call a “marketplace of agents.” Picture a hospital rather than a family doctor. Instead of one physician handling everything, you get a cardiologist, a radiologist, and a surgeon, each focused on a narrow job.
In the Princeton framework, sub-agents are each tailored to a specific task domain. They are developed separately, by different builders, and then work together on complex problems. No single agent has to be good at everything.
The guiding principle is straightforward. Specialization plus composition, according to the research, outperforms generalization at scale.
The headline number comes from SWE-bench, the evaluation framework the team used to measure results. SWE-bench is designed to test how well AI systems handle software engineering problems, which makes it a demanding proving ground for coding agents.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
An accuracy of around 70% on that kind of test is a meaningful result. It means the combined system succeeded on roughly seven out of every ten challenging tasks it was given.
Old ideas, new application #
The approach borrows from two familiar playbooks. The first is ensemble learning, the practice of combining multiple models so their collective judgment beats any individual one. The second is microservices architecture, where a large application is broken into small, independent services that each do one thing well and talk to each other.
What the Princeton team does is apply those ideas to autonomous agents built on language models. The paper also builds on earlier Princeton work, including SWE-agent, which focused on letting language models operate as software engineering agents. The new research extends that thread from a single agent to a coordinated cast of them.
What this means for the AI agent race #
The most immediate implication is architectural. If specialization and composition really do beat a single general agent, the question for builders changes. It stops being only “how do we make our model bigger” and becomes “how do we combine the right pieces.”
That opens a new set of design problems. Someone has to decide which agents get called, how they share context, and how their outputs are merged. The research puts those frameworks for combining diverse agents squarely on the agenda.
That caveat deserves attention. Benchmarks like SWE-bench are useful, but they are controlled environments. Real codebases are messier, and coordinating multiple agents adds its own failure points. If one specialist misreads a task, the error can travel down the chain.
There is also the question of accountability. When a single model gets something wrong, there is one system to inspect. When five agents from five different developers collaborate on a bad answer, figuring out what went wrong gets more complicated.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our