Agentic Systems as a Privacy-Preserving Search Problem

The Search Problem Framework

Jean-Denis Greze argues that the concept of "agent-to-agent" communication is less useful than viewing all LLM-based systems as a search problem. The objective is to ensure the context window contains the exact information required for a tool call or response, without human intervention. The ideal state is a single agent with access to all global data, but this is blocked by the "transaction cost" of privacy. Systems should be evaluated by how well they approximate this ideal while respecting security boundaries.

Strategies for Silo Navigation

To bridge the gap between private data silos and agent utility, five strategies are currently in use:

• Shared Trust Boundaries: Agents operate within a pre-defined group (e.g., an HR team) with access levels equivalent to a junior employee. This approach does not scale with model improvements because it still requires human oversight to manage the data.
• Privacy-Preserving Tools: Custom tools perform computations on private data and return only non-sensitive metadata or scores (e.g., returning a connection strength score rather than raw email content). This allows agents to act on insights without exposing the underlying private records.
• Sweeper Agents and Shared Silos: Agents automatically move information from private silos into shared spaces (like wikis or Airtable) based on defined policies. A "sweeper" agent can run daily to filter and migrate data, effectively desiloing information over time.
• Humans as Conduits: Agents ping humans to request permission to access or share information. This is inefficient at scale, as it often results in spamming irrelevant users for information they may not possess.
• Black-Box Search: An agent searches all silos automatically but only requests human approval at the final step of disclosure. The agent determines which specific user owns the necessary information and asks only that person for permission to share the result.

The Future of Auto-Mode

The most promising path for enterprise AI is the adoption of "auto-mode" for privacy policies. As LLMs become more capable, they will be trusted to enforce complex policies that determine what information is safe to share based on user roles and risk profiles. This allows the "low sensitivity" zone of accessible data to grow automatically as model capacity increases, creating a system that improves over time without requiring manual re-configuration.