Don’t Build Multi-Agents is a post on the blog of Cognition, the company behind Devin, that argues against a then-popular way of building AI agents: splitting a task among several subagents that work in parallel and combining their results. Its starting point is that building agents still lacks the kind of shared philosophy React gave web development, and that some libraries — it names OpenAI’s Swarm and Microsoft’s AutoGen — actively push multi-agent architectures, which the author considers the wrong way to build agents.
The post places reliability for long-running agents at the centre and names Context engineering as its core — which it describes as the next level of prompt engineering, doing the work of supplying a model with the right context automatically in a dynamic system, and effectively the most important job of engineers building AI agents. From a worked example it derives two principles, argues that architectures violating them should be ruled out by default, and recommends a single-threaded linear agent, optionally extended with a model that compresses the history of actions and conversation for very long tasks.
Key Points
- The post’s example is a “Flappy Bird clone” split into a background subtask and a bird subtask: one subagent misreads its subtask and builds a background in another game’s style, the other builds a bird that does not match, and the final agent is left combining two miscommunications.
- Copying the original task to each subagent is not enough, because in a production system the conversation is multi-turn and earlier tool calls and details may all bear on how the task should be interpreted.
- Principle 1: share context, and share full agent traces, not just individual messages.
- Even with shared context, parallel subagents that cannot see each other’s work make choices based on conflicting assumptions, giving inconsistent results — which the post states as Principle 2: actions carry implicit decisions, and conflicting decisions carry bad results.
- The simplest architecture that follows both principles is a single-threaded linear agent, whose context is continuous; its limit is context-window overflow on very large tasks.
- For truly long tasks the post presents one option: a separate LLM whose purpose is to compress the history of actions and conversation into key details, events and decisions — which it calls hard to get right, noting that Cognition has fine-tuned a smaller model for this.
- It describes Claude Code, as of June 2025, as spawning subtasks but never running them in parallel with the main agent and usually giving them only a question to answer rather than code to write, keeping the subagent’s investigative work out of the main agent’s history.
- It cites the Edit apply model pattern of 2024 as another design that suffered from miscommunication between models, and notes that edit decisions and their application are now more often made by a single model in one action.
- It judges that in 2025 agents cannot yet resolve disagreements through long-context dialogue with each other much more reliably than a single agent, so collaborating agents produce fragile systems with decisions too dispersed and context not shared thoroughly enough; the author expects this to improve as single-threaded agents get better at communicating with humans.
Context
The post is written from the perspective of a company building its own agent product, and presents its principles as ones Cognition’s team keeps relearning and builds its internal tools and frameworks around. It reports no measurements, and the author expects the field’s standards to change, calling for some flexibility and humility.