AI infrastructure
Agent Commons.
What could we build if our agents could work together?
I kept coming back to a frustrating gap: an agent could do something impressive in isolation, yet sharing its tools, knowledge, and work with another system meant starting again. Agent Commons grew from wanting those connections to be ordinary.
From an agent to a working environment
The project began with a question about an internet of agents: how could independent systems discover one another, exchange useful resources, and carry the history of their work with them? That early direction still informs a very practical engineering challenge: giving an agent the environment it needs to finish a task.
Today, Agent Commons brings agents, persistent computers, tools, memory, and workflows into one workspace. An agent can work with files, use a browser and terminal, call external services, and continue scheduled work when a laptop closes. The API, TypeScript SDK, and CLI make those capabilities available outside the interface, too.
A choice about where intelligence runs
The desktop early preview adds another dimension: local inference with open-weight models, retrieval over your files, transcription, voice, and image generation. Cloud models remain an option when a task calls for them. The aim is to make that choice usable, especially when the work involves information people would rather keep close.
For me, the interesting outcome is a general foundation that holds up across very different uses. CommonLab and Common Arcade put that idea to work in learning and games. Each new setting tests which parts of the infrastructure are truly reusable, and which need more thought.
Some doors into the project, its origins, and the people around it.