@zostaff: Walden Yan, co-founder of Cognition (the team behind Devin), in an AI Engineer talk on agent harnesses: "A lot of the m…
Summary
In an AI Engineer talk, Cognition co-founder Walden Yan recommends against multi-agent systems, saying they are fragile and that the team behind Devin now uses a single agent with coherent context.
View Cached Full Text
Cached at: 08/09/26, 07:14 AM
Walden Yan, co-founder of Cognition (the team behind Devin), in an AI Engineer talk on agent harnesses:
“A lot of the multi-agent systems we’ve played with can be quite fragile. We’ve ultimately moved away from them.”
Not an interview. The man who builds Devin, telling a room of agent builders to stop doing the thing everyone is racing to build.
You picture an AI company: one agent per role, all collaborating, a whole org replaced. That is the demo everyone chases.
The team that ships real coding agents killed it. They run a single agent, because two agents that can’t see each other’s context make conflicting decisions and quietly break.
The harness that wins is not a swarm. It is one agent that never loses the thread. Multi-agent looks powerful and shatters at scale. One coherent context beats ten clever ones that can’t talk.
Similar Articles
A developer shares insights on how to maximize AI agent capabilities, arguing that simpler setups and understanding core principles are more effective than complex harnesses and libraries.
A developer shares insights on how to maximize AI agent capabilities, arguing that simpler setups and understanding core principles are more effective than complex harnesses and libraries.
@SuJinyan6: https://x.com/SuJinyan6/status/2073955240349770069
This blog post by SuJinyan6 examines the evolution of AI agents from simple LLM+tool use to context engineering and long-running harnesses, citing Anthropic's recent work and discussing how agent capability is now a system-level property involving multiple components.
@hwchase17: https://x.com/hwchase17/status/2053157547985834227
The article outlines a systematic 'Agent Development Lifecycle' (Build, Test, Deploy, Monitor) for creating and managing AI agents effectively, highlighting key frameworks like LangChain, LangGraph, and CrewAI.
@DerekNee: everyone is talking about agent loops, harnesses, and self-evolving agents. but almost no one is talking about the actu…
The author argues that a single giant agent is ineffective for running a company, and describes their approach with Matrix, an operating system for autonomous work that organizes agents into a workspace brain, department leads, and scoped workers with proof loops.
@NainsiDwiv50980: Everyone is talking about AI agents. Very few people are building the thing that actually makes them powerful: Context.…
A Twitter thread argues that the key to powerful AI agents is not better prompts but accumulated personal context and memory systems, highlighting Obsidian as a tool for compounding knowledge. The author predicts a widening gap between those using AI alone and those combining it with personal context.