How Many AI Agents Can One Programmer Run at Once?

Ryo Lu says the best programmers he knows can maybe run only four agents at once. A person looks at one chat thread at a time and cannot take in that much information at once. The problem that follows is how to help someone who would otherwise be chatting with 15 agents at once or reviewing every change those agents make.

Cursor's interface started as manual coding, files and lines typed by hand. Completion came next: the product finished the next function, jumped to the right file, and tried to keep the person in flow while they typed. Chat made the model aware of the whole codebase and able to answer questions. Editing let it apply changes, and tools made the work more agentic. People are still the limit on how much of that can be watched.

As agents get more autonomous, with more context and more tools, and a clearer sense of what someone wants, a captured intent and a plan can turn into a lot of work, and the person becomes the bottleneck. The next interface sits higher. Instead of touching the code directly, you chat with it. Instead of chatting with each agent, you plan with it and look at a dashboard of the plans and tasks. Designers may want that layer to be visual and directly manipulable: click something, make it bigger, and it is bigger. Soleio asked whether that layer unlocks a concentric circle of new users, and Ryo Lu agreed.