When More Agents Pay Off and When They Don't
mlj.solutions
mlj.solutions Víc AI agentů neznamená lepší výsledek. Když si úkol předává několik agentů za sebou, každý dostane jen zkrácenou verzi. S každým předáním se může ztratit informace, na které rozhodnutí stojí. Pro firmu to často znamená, že jednodušší řešení funguje líp i levněji. Pokud řešíte, jaké nastavení dává smysl u vás, odkaz na konzultaci najdete v biu!
View on InstagramWhat the study is about
More AI agents don't necessarily mean a better result. The study shows that the reason is the passing of only condensed messages between agents, which can save unnecessary context but can also lose important information. The authors describe this as an information bottleneck and tested it in controlled experiments.
Key numbers
18
controlled experiments
The authors tested the theory in controlled experiments comparing a single agent with a team of agents.
5
benchmarks
The tests covered five different benchmarks so that the conclusions are not tied to a single type of task.
3
model sizes
Results differed depending on model capability, so the authors tested both weaker and stronger models.
What this means
- More agents don't mean a better result, the benefit depends on how much useful information is lost when it is passed between agents.
- A team helps weaker models more, if the messages between agents are sufficient, it pays off mainly with weaker models.
- Strong models gain little, they can extract useful information even from a longer context, so condensing it brings them little and the gain can even reverse.
- Unrestricted handoffs don't beat a single agent, a team of agents can imitate a single one if they can pass on the full context; the advantage appears only with limited handoffs.
- Agent design is a matter of trade-offs, the question is how much information to compress so that space is saved without losing what is essential.
Our take
More agents don't automatically mean a better solution. That's why, for every process, we first verify whether passing information between agents pays off, and only then design the architecture.
Zdroj
Yu, W., Zhou, L., Dong, X., Barath, S. S., Staunton, D., Yoon, B.-J., Qian, X., Caverlee, J., & Ji, S. (2026). When Do Multi-Agent Systems Help? An Information Bottleneck Perspective. arXiv:2607.16133. https://arxiv.org/abs/2607.16133
Don't miss these posts
AI Remembers a Claim Even When You Tell It It's False
When a model is fine-tuned on documents that deny a claim, it often learns the claim as true anyway. The authors named this effect Negation Neglect.
The Skill That Explains Anything to Anyone
ELI5 is an open-source skill for Claude Code that adapts explanations to the audience. With the skill it passed 10 of 12 criteria, without it only 5.
A Correct Answer Is Not the Same as a Verified Analysis
TrustDABench tests whether AI working over tables can refuse to answer when the data isn't sufficient. The best model reached only 24.21% in reliability.
Want to implement AI in your company?
We'll go through your needs together and propose an approach that delivers measurable results.
Write to us
