Agents' next frontier: agent-to-agent and network effects — Jean-Denis Greze, Town

Overview

• Speaker: Jean-Denis Greze
• Channel: AI Engineer
• Main topic: Agent-to-agent communication, network effects, and privacy-preserving strategies for LLM agents
• Purpose: To explore how multi-agent systems and agent-to-agent network effects work, highlight the privacy-compute tradeoffs, and propose architectural strategies for future AI agent ecosystems. Jean-Denis Greze, CTO at Town, discusses the next frontier in AI engineering: agent-to-agent interactions and network effects. He reframes running agents as a search problem over context windows and outlines five key strategies for feeding the right data into LLMs while balancing power, privacy, and security in multi-agent systems.

Topic Map

Running an agent as a search problem (01:12)

• Explanation: Reframing LLM agent execution as a search problem where the core goal is ensuring the context window has the right information right before a tool call or user response.
• Key claims:
  • Most LLM systems are fundamentally search problems.
  • Agentic search involves giving an agent tools to search through content spaces to populate the context window.
• Examples:
  • Human-populated context windows versus automated agentic search tools.
• Terminology:
  • context window
  • agentic search
  • tool call
• Why it matters: Shifts the paradigm of agent engineering from just prompting to optimizing context population across systems.

Five strategies for LLM data access and trust boundaries (04:12)

• Explanation: Five strategies used by companies to get the right data into LLM context windows within trust boundaries.
• Key claims:
  • Approximating access to everything within trust boundaries is popular for IT and security teams but fails to scale or de-silo.
  • Privacy-preserving cross-silo tools trade off power and privacy.
• Examples:
  • Relationship strength tool that returns a score instead of raw email content.
  • Shared wikis and Airtables as shared silos.
  • Human-in-the-loop and black-box compute approaches.
• Terminology:
  • trust boundaries
  • cross-silo tools
  • shared silos
  • human-in-the-loop
  • black-box compute
• Why it matters: Provides concrete architectural patterns for solving privacy and data access challenges in enterprise multi-agent applications.

Key Points

Agent-to-agent interactions unlock massive network effects

• Explanation: Just as human work relies on collaborating with other people, AI agents working together create exponential network effects and value.
• Evidence: Real-world work models demonstrate that more is better when agents can collaborate across boundaries.
• Practical implication: Engineers must build systems that allow agents to seamlessly communicate and share data safely.

The Coase Theorem applied to agent silos

• Explanation: Even with infinite context windows, privacy and security constraints prevent humans from sharing all context, creating silos.
• Evidence: Economic principles show transaction costs and privacy concerns limit ideal information sharing.
• Practical implication: Multi-agent systems must navigate privacy constraints rather than assuming complete data access.

The frontier is automated policy enforcement and auto-mode

• Explanation: As models and safety policies improve, systems will increasingly automate decisions around data sharing and human oversight.
• Evidence: Coding and approval workflows are moving from strict manual checks to autonomous handling with low-sensitivity zones.
• Practical implication: Developers should define low-sensitivity zones and trust boundaries to allow systems to scale autonomously over time.

Frameworks, Models & Processes

Five Strategies to Get LLMs the Right Data

• How it works: A spectrum of approaches ranging from broad access approximations to automated black-box compute with human approval.
• Components:
  • Approximate access to everything within trust boundaries
  • Privacy-preserving cross-silo tools
  • Create shared silos
  • Humans as the conduit
  • Black-box compute with approval on inputs
• When to use: When designing enterprise AI agent architectures and determining data governance policies.

Examples & Case Studies

A tool checks everyone's Gmail across a company and returns a relationship strength score without exposing raw conversation content.

• Illustrates: Privacy-preserving cross-silo tooling (signal, not content).
• Lesson: You can enable agentic discovery across company silos while preserving user privacy by returning computed metrics instead of raw text.

Actionable Takeaways

• Immediate:
  • Evaluate whether your agent architecture relies on manual data silos or automated discovery.
  • Design tools that separate signal from content to preserve privacy.
• Strategic:
  • Anticipate a shift toward automated ('auto-mode') policy enforcement in multi-agent systems.
  • Recognize that multi-agent network effects will drive the next wave of AI utility.
• Questions to investigate:
  • How do we prevent prompt injection across siloed agent networks?
  • Can we trust autonomous agents to make privacy disclosure decisions correctly?

Claims Worth Verifying

• Running an LLM agent is fundamentally a search problem focused on context window optimization. (conceptual)
• Black-box compute architectures with automated input verification can eliminate the need for manual human-in-the-loop conduits. (architectural)

Notable Quotes

"I don't think agent to agent makes much sense as a concept. So I want to reframe the entire talk in terms of search." (at 1:52) "If this is only strategy, you lose over the next 3 years." (at 7:02)

Compressed Summary

• Agents can be framed as search problems over context windows.
• Multi-agent systems face data silo and privacy challenges modeled by economic theory.
• Five strategies exist for data access: trust boundaries, cross-silo tools, shared silos, human conduits, and black-box compute.
• The future trend is moving toward autonomous ('auto-mode') policy enforcement with human review.
• Keywords: agents, search, privacy, silos, network-effects
• Core insight: Multi-agent systems unlock massive network effects, but scaling them requires sophisticated privacy-preserving cross-silo tools and automated policy enforcement rather than naive shared context.