Multi-Agent Systems as a Search Problem
Jean-Denis Greze argues that the concept of "agent-to-agent" communication is often a distraction. Instead, LLM systems should be viewed as a search problem: the goal is to ensure the right information is present in the context window at the exact moment a tool call or response is generated. The ideal state is a single agent with universal access to all data, but this is blocked by the "transaction cost" of privacy. The effectiveness of any multi-agent system is measured by how well it approximates this ideal without violating security boundaries.
Five Strategies for Cross-Silo Data Access
Greze evaluates five approaches to navigating data silos, ranging from manual to automated:
- Shared Trust Boundaries: Agents operate within a defined group (e.g., an HR team) with access equivalent to a human member. While popular, this does not scale with model improvements because it remains a manual, siloed process.
- Privacy-Preserving Tools: Custom tools that perform a specific task (e.g., calculating a relationship score between employees and external contacts) and return only the result, rather than raw data. This trades power for privacy but requires manual tool creation.
- Shared Silos & Sweeper Agents: Creating a shared space (like a wiki or database) where information accumulates. A "sweeper" agent monitors private silos and automatically migrates non-sensitive information to the public space based on defined policies. This is identified as the most promising near-term strategy.
- Humans as Conduits: Using humans to approve data sharing requests. This is inefficient at scale, as it often results in spamming users with approval requests for low-probability information matches.
- The Black Box Approach: An agent with broad access that performs searches across silos automatically. It only requests human approval at the final step, when it has identified the specific information needed and the owner of that information. This requires high trust in the agent's ability to filter data correctly.
The Future: Auto-Mode and Policy Enforcement
The frontier of agentic systems lies in "auto-mode" for privacy. As models become more capable, they will better handle policy enforcement, allowing for more dynamic, automated disclosure of information. The key to scaling is defining a "low-sensitivity zone" where the agent is trusted to make decisions. Over time, as policies are better encoded and models improve, this zone will expand. The ultimate, high-value frontier is cross-company agent collaboration, where organizations agree to shared agent protocols to facilitate mutually beneficial data exchange (e.g., in finance or lending).