Notes and posts by @spolu
Agents need more than local computers or sandboxes: they need shared servers where humans and agents can collaborate over durable state. That shift points toward hyper-personalized business software, with authorization as the hard part.
We standardized on the word "harness" to describe agent environments, but a harness is inherently single-player. Three years ago I predicted we'd get the DOOM, Civ, and Counter Strike of LLMs. We're getting closer — but multiplayer modes are still far away.
Great writing may be the last refuge of human superiority on models. It's quite surprising to me that we still detect fairly well AI slop vs great writing but have all, as software engineers, accepted AI generated code as the new norm.
There's been a lot of discussion about MCP lately. We're in a "long live the CLI" moment. Some of the criticism is deserved but it's also important to recognize what MCP gets right.
Coding agents are just better than us at most well-defined, well-packaged coding tasks. It is increasingly rare to find coding tasks where doing it manually will be more efficient than packaging the task clearly for a set of agents to tackle it.
Most AI use out there is single agent, single player. Our hypothesis: collaboration is on the critical path of accelerating companies with AI.