A new study says stacking more AI agents on a task often makes things worse, not better.
Researchers set out to map exactly when multi-agent collaboration beats a single, well-built AI agent, rather than assuming it always does. They found multi-agent setups pay off specifically on long, drawn-out tasks where the pieces do not depend heavily on each other, while tightly linked, sequential work is still handled better by a single agent. Based on that split, the team built SAIGE, a lightweight collaboration method that spawns agent instances only when needed and connects them through a graph based on what content they actually share, instead of a fixed team structure. In tests on long, complex benchmark tasks, SAIGE held task performance while using less context, and simply adding more agents or letting them recurse deeper did not reliably improve results.
That matters because a lot of current agent tooling treats more agents, more delegation, and more orchestration as an automatic upgrade. This research says the opposite: multi-agent value is narrow and shaped by task structure, not a universal replacement for a good single-agent harness. Teams building agent products get a concrete signal here: check whether a task is long-horizon and loosely coupled before reaching for a multi-agent architecture at all.
It is a useful reality check in a year when adding more agents has become the default answer to almost every AI product problem.