Multi-agent AI's real problem is privacy governance, not model power

The speaker frames agent-to-agent AI as a search problem: getting the right information into the context window before a tool call. The ideal is a single agent with access to all data, but privacy and security prevent that. Instead, teams must approximate this ideal through five strategies: trust-boundary agents, custom privacy-preserving tools, shared silos,

humans-in-the-loop, and a black-box approach where an LLM accesses all data but only requests human approval for specific disclosures. The most immediately practical strategy is the AI-maintained shared wiki, where an LLM automatically surfaces private information into public spaces based on policy. This works best in small, high-trust companies. The black-box approach is

more powerful but requires trust in the LLM's judgment and auditing capabilities. The speaker argues that as models improve, systems should be designed with a 'low sensitivity zone' where the LLM can make autonomous decisions, expanding this zone over time. The real challenge is organizational governance, not model capability. Teams must decide what

data is safe to share, how to audit decisions, and how to handle prompt injection. The speaker predicts that agent-to-agent data sharing will first succeed within companies, then eventually across companies in specific high-trust scenarios like finance. The key is to start with low-risk, high-value use cases and scale trust as models improve.