Summary
Jean-Denis Greze, CTO of Town and former seven-year CTO at Plaid, argues that multi-agent AI architectures are fundamentally a search problem rather than a coordination problem. The central thesis: the theoretically optimal AI system is one with perfect access to all relevant information in a single context window — multi-agent systems are simply practical approximations of that ideal, constrained by privacy, security, and Coase theorem-style transaction costs.
The talk walks through five concrete strategies practitioners use to get the right data into an LLM’s context at call time: approximate access within trust boundaries (e.g., a shared agent across spouses or small teams), agent-to-agent delegation for routing to specialists, shared silo structures like company wikis and skills repositories, and a “sweeper AI” concept — an agent that monitors private silos and surfaces appropriate information into shared spaces on a schedule, requiring either human approval or policy-based rules to govern what flows out.
Greze draws on real-world experience building consumer-facing agents at Town to explain why each strategy has meaningful limits — trust model complexity, human approval friction, and the difficulty of encoding privacy policy as LLM instructions. The framework is useful for any engineering team designing systems where multiple agents must collaborate across organizational or personal data boundaries without violating access controls.
📺 Source: AI Engineer · Published September 03, 2026
🏷️ Format: Deep Dive
