More agents don't coordinate themselves. You have to design for it.
Anthropic's research into multi-agent systems found that stronger models don't produce better teamwork on their own — coordination has to be engineered in, the same way access control and audit trails are.
More agents isn't more capability — it's a new failure mode
The instinct once a single agent works is to add more of them. Anthropic's research is a useful check on that instinct: agents facing the same situation tend to behave far more similarly to each other than a team of people would, and that similarity is exactly what breaks systems at scale. A dozen agents converging on the same shortcut, the same resource, the same assumption isn't a coordination win — it's a pile-up. Capacity doesn't compound if every agent reaches for the same lever at once.
Agents don't arrive with a reputation, so the system has to supply one
A team of experienced analysts develops a feel for which colleague's read to trust on which topic. Agents don't have that history, and left alone they don't compensate well — an unreliable input can drag the group toward the wrong conclusion as easily as a reliable one, especially when the accurate read comes from a single dissenting source against an apparent consensus. If multiple agents are feeding conclusions into a decision, something in the system has to track provenance and confidence explicitly, because the agents won't build that judgement on their own.
Design the conflict resolution before the agents meet
Give agents overlapping goals and no shared referee, and the failure mode isn't paralysis — it's escalation. Left to resolve conflicts themselves, agents have been observed working around each other, then locking each other out of shared resources entirely. The lesson for anyone deploying more than one agent against shared systems: decide up front what's reversible, what requires a shared arbiter, and what a human has to see before it happens — don't assume the agents will negotiate it out.
Coordination doesn't emerge from smarter models. It has to be designed in — the same way access, audit, and rollback are.
Source perspective: Anthropic Research ↗. This AventeqAI brief is an original synthesis and commentary, not a reproduction of the source article.