Korva is a scroll of the internet’s best ideas.
Agents coordinate best with minimal structure
Why use little structure
The usual design puts a coordinator on top that hands tasks to child agents, which work alone and return answers. That breaks down in small ways. Two children with similar tasks usually cannot talk to each other, and a child with a clarifying question must either return without solving the problem or guess what the parent wanted.
How agents talk
Patching those gaps adds complexity to the scaffold, so OpenAI baked in as little structure as it could and let the agents work out coordination themselves. The result resembles people collaborating over Slack. One agent announces an answer, another reports a different one, and they trade explanations of how each arrived there until one concedes and broadcasts the change to the rest.
What was hard
Getting there was hard. Agents tend to collapse into each solving the problem independently, and a message arriving interrupts a model's deep chain of thought. Stronger, more general models made the behavior easier to develop.
How do most multi-agent systems split work?
The common pattern is orchestrator and workers: a lead agent breaks a task into pieces, sends each to a sub-agent, and combines what comes back. Anthropic described this design for its Research feature in June 2025. Sub-agents search in parallel, each in its own context window, and return summaries. Anthropic reported that this setup beat a single-agent baseline by 90.2 percent on its internal research evaluation.
Why does an interruption disrupt a model's thinking?
A language model writes one token at a time, and each token depends on all the text before it. A message inserted mid-reasoning becomes part of that text, so everything the model writes afterward is conditioned on it. A person can finish a thought before reading a notification, but a model has no separate place to set one aside. The transformer architecture behind these models, introduced in a 2017 Google paper, lets every new token attend to everything earlier in the context window.
Why does adding collaborators add overhead?
Fred Brooks argued in his 1975 book The Mythical Man-Month that adding people to a late software project makes it later. Part of the reason is that possible communication links grow faster than headcount: n people have n times n minus one, divided by two, pairs. Five people have ten pairs, and fifty have 1,225. Brooks drew on his experience managing IBM's OS/360 project.
More from Dwarkesh Podcast
