
The Deep View: Conversations #37 - The race to control AI agents begins - James Everingham
Apr 9, 2026
James Everingham, CEO of Guild AI and former Meta developer-infrastructure lead, builds a control plane to govern AI agents across enterprises. He discusses agent supervision, viral internal adoption that proved product-market fit, why safety infrastructure is a competitive moat, and how workflow-specific agents, open-source plays, and cost-control strategies shape the future of agent deployment.
AI Snips
Chapters
Transcript
Episode notes
Meta's Internal App Store Sparked Viral Agent Growth
- At Meta James helped build an internal managed software center where engineers could browse, fork, and deploy agents, which sparked viral internal adoption.
- That internal “app store” multiplied agents quickly and forced the team to find tens of thousands of developer servers to run instances.
Deterministic Rails For Non‑Deterministic Agents
- Non-deterministic agents require a deterministic supervisory layer to contain unpredictable behavior.
- James recommends evaluating agents with repeated task suites, intercepting data/execution access, and enforcing policies via the control plane.
Harnesses Run Inside Control Planes
- A harness is the code pattern that controls a model; a control plane manages and validates harness actions across the system.
- James compares it to OS protection layers: harnesses run inside the control plane which enforces policies and records activity.

