We keep trying to make AI agents smarter by giving them more: more tools, more files, more instructions, more memory.
But perhaps the real advantage comes from giving them less.
I recently used Contessa to create a workspace for Pi, a coding agent. Instead of exposing my entire machine, I described the world Pi needed in YAML: one repository, two relevant skills, and a private directory for its session history.
Contessa projected that selection as an ordinary filesystem. Pi could still use familiar tools—shell commands, search, compilers, Git—but irrelevant projects, credentials, extensions, and old sessions were absent.
Nothing about the model changed. Its environment did.
This matters because an agent’s context is not just what appears in its prompt. It is everything the agent can discover and act upon. A carefully constructed filesystem can become a practical boundary around that world.
The interesting question may no longer be:
How much context can we fit into the model?
It may be:
What is the smallest world in which the agent can succeed?
That is the idea behind Contessa: curate the context, then serve it as a filesystem.