How and When to Build Multi-Agent Systems

The debate it settles

In June 2025 two posts seemed to disagree. Cognition, the team behind the Devin coding agent, published "Don't Build Multi-Agents", arguing that splitting work across agents loses context and produces conflicting decisions. Anthropic published how it built a multi-agent research system that outperformed a single agent on broad research questions. This post, published on 16 June 2025, argues they largely agree once you look at the kind of task each was describing.

The key ideas

• Context engineering is the real job. The hardest part of building agents is getting the right information into each model call at the right time. Multi-agent systems make that harder, because every sub-agent needs enough context to do its piece. Anthropic found that vague delegation such as "research the semiconductor shortage" led sub-agents to duplicate each other's work.
• Reading parallelises; writing does not. Gathering information from many independent directions splits naturally across agents. Producing one coherent output (code, a document, a decision) goes wrong when several agents each make choices the others do not know about. Anthropic's system uses many agents to research, then one agent to write the final answer.
• Multi-agent pays off for broad, valuable tasks. It suits breadth-first questions with independent directions, where the extra cost and tokens are justified. It does poorly where agents need shared context or depend on each other heavily.
• Control matters. The post argues for orchestration tools, such as LangGraph, that let developers see and control exactly what goes into each model call, rather than frameworks with hidden prompts.

How to use it in your own design

Before splitting a system into agents, ask:

1. Is the work mostly gathering information from independent sources? Multiple agents may help.
2. Does it produce one artefact where decisions must stay consistent? Keep one agent in charge of writing.
3. Can you clearly specify each sub-task's goal, output format and boundaries? If not, delegation will waste tokens.
4. Is the task valuable enough to justify several times the cost and latency?

If most answers point to a single agent, improve its tools and context first.

Read alongside

The post is much more useful if you read the two pieces it responds to: Cognition's "Don't Build Multi-Agents" and Anthropic's "How we built our multi-agent research system". Then Microsoft's AI agent orchestration patterns guide gives names and trade-offs for the arrangements you might choose.

It is a vendor blog, and it naturally positions LangGraph as the answer. The reasoning about tasks and context holds regardless of framework.

Common questions

How long is it? A ten to fifteen minute read; allow an hour with the two linked posts.

Is it still relevant? Models have improved since mid-2025, but the trade-off between parallel reading and consistent writing is structural and still applies.

Best for: engineers designing agent systems who want a clear rule of thumb for when multiple agents are worth it.