Self-Optimization: the reinforcement loop and the future of Agent Trust
About this session
Discover how to build self-improving AI systems through a continuous reinforcement loop that identifies hidden failures, proposes fixes for human approval, and scales across fleets of AI agents. Learn why connecting telemetry, evaluations, and code context is essential to diagnosing issues accurately, improving agent performance, and building greater trust in production AI.
Speaker
Key takeaways
- Self-improving AI is infrastructure you build, not a switch you flip. It rests on three parts: a blueprint, labor to run it at scale, and a foundation of good signals.
- The Reinforcement Loop is the blueprint. It reads real production conversations, catches failures (including ones that look healthy) and drafts fixes for a human to approve. Two customer examples show the kinds of failures offline evals miss: a betting agent that gave opposite answers to the same question, and a fintech agent that reported zero errors while redoing half its work.
- An Agent Fleet is the labor. Once you have hundreds or thousands of agents, no team can run that loop by hand. Specialized agents for monitoring, triage, troubleshooting, remediation, tuning and more do it at scale.