Jean-Denis Grèze is cofounder and CEO of Town, which builds personalized AI assistants. Previously CTO at Plaid and an engineering director at Dropbox, he writes and speaks about engineering ownership and agents working across information boundaries.[1][3]
Start with Agents’ next frontier: agent-to-agent and network effects →
Education
Links
X profile Personal writing directory Engineering ownership interview Town product Cofounder and CEO · Town
Career
Grèze’s first-person interview traces undergraduate computer science, an unfinished Ph.D. and a detour into legal practice before returning to engineering. At Dropbox and Plaid, he emphasized solving business problems within real resource constraints. His ownership model gives teams closest to a problem room to decide across functions, while leaders handle staffing, budgets and competing priorities.[3]
Contributions
Town presents an assistant connected to existing work tools and recurring routines. It accumulates context about a person’s work so that assistance can extend across email, scheduling and other applications. That product focus connects Grèze’s organizational interest in collaboration with software that carries context between everyday tasks.[5]
Ideas
His agent-collaboration talk compares shared agents, narrowly scoped tools and shared knowledge stores. A tool can return a relationship score instead of disclosing the underlying private email; a proposed sweeper agent can move approved low-sensitivity material into a shared wiki. The challenge is supplying the next model call with useful context while controlling what information may leave each person’s private workspace.[4]
Tools, writing & talks
Writing, research & talks
Start here
A comparison of architectures for sharing useful context between agents while preserving information boundaries.
Start with the relationship-search example and identify what each proposed architecture reveals to another person. Familiarity with agents helps; the talk’s proposed sharing systems need disclosure controls as well as retrieval quality.
Presenter[4]Town's CTO rejects agent-to-agent as a framing, calling every LLM system a search problem where privacy is the transaction cost blocking one agent with access to everything. He then grades five approximations: shared trust boundaries, privacy-preserving custom tools, sweeper-fed silos, human conduit
Catalog workOriginal source