Exo: Let Agents See Their Own Code and Logs

Why it matters

Alex Krentsel from UC Berkeley and Google Research joins Latent Space to discuss Exo: the idea that AI agent harnesses should expose their own code, logs, and tool definitions to the agent itself. This transparency enables self-debugging and continuous improvement, offering a fresh perspective on autonomous, introspective systems beyond the traditional black box.

Key takeaways

• The harness acts as an agent's operating system, yet remains a black box; Exo proposes exposing it to the model.
• By letting agents read their own logs and code, they can identify and fix errors autonomously, reducing human intervention.
• Transparency is built around composability: tool descriptions, Python environments, and execution logs are all queryable, creating a feedback loop.
• Exo targets long-horizon autonomy, aiming for agents that understand and improve their own runtime like a developer would.