Don’t Build Multi-Agents
Read date: 21 Sep 2025
The Cognition team doesn’t argue that multi-agent systems are “bad,” but rather that they tend to be ineffective when sub-agents run in parallel without proper coordination. That’s where the main issues arise: each agent can make decisions misaligned with the others, leading to an inconsistent final result.
That’s why they highlight two core principles:
-
The full context should be shared across all agents, so they don’t work “blind.”
-
Sequential coordination is much more effective than disordered parallelism.
But here’s an important point: they don’t recommend multi-agents by default. Their central thesis is that context engineering—structuring and maintaining context clearly and effectively—is usually more powerful than multiplying agents. In many cases, a single well-designed agent with strong context handling will do the job as well or even better.
When a multi-agent setup does make sense, they propose running sub-agents sequentially. And if the task requires long runtimes, they suggest adding an LLM to summarize the context along the way, so that later agents can operate with a condensed and coherent view of the conversation.
In short:
-
Multi-agents are not the enemy, but uncontrolled parallelism is.
-
What matters isn’t the number of agents, but the quality of the context.
-
Start simple: one agent with strong context. Scale to multi-agent only when strictly necessary.