Jean-Denis Greze

Jean-Denis Greze is co-founder and chief executive of Town, which builds personal AI assistants that use accumulated context to help people across their everyday tools. Previously CTO at Plaid and an engineering leader at Dropbox, he approaches agent collaboration as an information problem: how to give a model the context it needs to act effectively when that information belongs to different people and cannot freely be shared.

From law to engineering leadership

Greze studied computer science at Columbia University, earned a law degree at Harvard Law School, and practiced law before returning to technology through work with vacation-rental businesses. He also built software for hedge funds. At Dropbox, he progressed from individual contributor to engineering director before joining Plaid. His account of that career and his engineering values connects those transitions with his approach to building teams.

At Plaid, Greze advocated engineering ownership: teams closest to a problem should have room to make decisions, with leaders intervening when staffing, budgets, or competing priorities require resource tradeoffs. His leadership approach also called for constructive disagreement across functions. Trusting another department should not prevent engineers from questioning a decision or helping work through a problem. Ownership requires both autonomy and a willingness to engage beyond one's immediate responsibilities.

Town’s pivot to personal assistance

Greze founded Town with Tony Vincent in late 2024; the two had worked together at Dropbox. The company initially pursued AI-assisted tax preparation, then shifted toward personal assistance. Administrative work surrounding the tax effort—arranging customer calls, answering recurring questions, and recovering context before meetings—helped reveal that opportunity. Their account of Town's development describes the difficulty of finding a workable go-to-market approach for the original product and the move toward helping people manage everyday work.

Town's assistant uses an accumulating understanding of a user's voice, relationships, priorities, and judgment to help across existing tools. The product aims to make assistance useful through ordinary activity rather than require users to design elaborate workflows. The company announced a $55 million Series A led by Andreessen Horowitz on June 3, 2026.

Greze treats greater autonomy as something an assistant must earn. Town's initial email automations create drafts: users can correct a poor draft before it reaches someone else, while an incorrectly sent message can damage their reputation. Its approach to granting responsibility uses repeated approvals and corrections to assess whether a workflow is ready for automation, with users retaining control over permission changes.

Getting useful context across information silos

In his talk on agent collaboration and network effects, where he introduced himself as Town's CTO, Greze develops the information problem behind that product ambition. He argues that the decisive step in many LLM systems is getting the right information into context immediately before an answer or tool call. He uses an imagined agent with access to all relevant information as a benchmark for collaboration. Real systems cannot grant that access indiscriminately because privacy and security constrain what people may see. The useful question is how closely a system can approximate that informational advantage while respecting those constraints.

• Shared agents: Greze compares five ways to work across information silos. A shared agent can operate within an existing trust group, such as an HR team, using the information its members are permitted to access. This fits familiar organizational access controls, but he argues that it also creates another silo whose scope people must define and maintain. His test for an architecture is whether it needs less human coordination over time and becomes more capable as models improve.
• Purpose-built tools: Purpose-built tools offer another approach: disclose a useful result without exposing the underlying records. To find a colleague who knows a prospective customer's finance executive, a tool could inspect email data and return relationship-strength scores while keeping conversations private. An agent could then ask the best-connected colleague for an introduction. Greze says Town uses tools built around this kind of privacy tradeoff for recurring user needs. Their limitation is that people still have to design each tool and explain what access and disclosure users are accepting.

AI-maintained shared knowledge

His strongest near-term bet is AI-maintained shared knowledge. Agents can contribute reusable information to a common wiki, database, or repository so colleagues do not repeatedly rediscover it. A shared database-profiling instruction illustrates the benefit: one engineer improves the method, and another engineer's agent can use it when diagnosing a slow query. The shared resource lets useful work carry forward across people and tasks.

Greze proposes a “sweeper” agent inside each private information silo. It would review new material, consult a policy describing what must remain private, and move appropriate information into shared company spaces. One version asks a person to approve proposed contributions; another lets the model enforce the disclosure policy itself. He expects smaller, high-trust companies to adopt automated sharing sooner, particularly where sensitive categories such as finance and HR information can be clearly identified. This is a proposal and forecast, rather than a claim that such systems have already solved disclosure reliably.

Permission, protected search, and disclosure risks

The remaining approaches differ in when they involve people. An agent can request permission to search each person's information and then request permission to share its findings. For a question only a few employees can answer, however, this can interrupt an entire company. Greze explores a more ambitious protected-search design: the system searches across private silos, identifies the information actually needed, and asks its owner for approval at the moment of disclosure. In the introduction example, it would approach the best-connected colleague instead of asking everyone to participate. He says he has seen little practical use of this approach, and it requires substantial trust in the system's access and disclosure controls.

Even revealing a relationship can expose something sensitive. Greze gives the example of a question about recruiter contacts that could reveal an employee's job search. He also identifies prompt injection, mistaken disclosures, and persistent errors in shared knowledge as hazards. His personal wiki continued using an agent's old name after he renamed it; a similarly persistent mistake about a business could have more serious consequences. Approval rules, logging, reversibility, and audit access therefore remain part of the design problem, including for a system described as a black box.

Greze recommends beginning with a clearly defined category of low-sensitivity information that a model may share automatically, while retaining human review for more sensitive decisions. His expectation is that better models and better policies will gradually expand what can safely be automated. Collaboration across companies could create further network effects where organizations benefit from combining otherwise private information, but he leaves the workable use cases and trust arrangements unresolved. He expects agents to make more privacy decisions while openly questioning whether he trusts a future in which they make all of them.