Summary
Soheil Feizi, founder and CEO of RELAI and associate professor of computer science at the University of Maryland, presents a rigorous framework for continual learning in AI agents — the process of converting real-world production failures into durable, verifiable improvements without triggering regressions in previously working behavior.
The talk identifies two core challenges: obtaining meaningful feedback from live systems (where there are no structured benchmarks, only raw logs and implicit user signals), and acting on that feedback in a provably safe way. Feizi surveys current approaches across three agent layers — model weights, harness (prompts, skills, tools, code), and memory — reviewing methods including GEPPO, prompt search via evolutionary algorithms, LETA, and MemZero. Each approach has distinct tradeoffs around testability, scalability, and risk of silent regression.
The centerpiece is a concept Feizi calls “verifiable continual learning”: a three-step framework requiring a replayable executable test derived from the failure, a measured delta showing the fix actually helps on that test, and a regression suite confirming nothing previously passing is broken. RELAI’s tooling at rely.ai implements this pipeline end-to-end — converting session logs and human expert feedback into synthetic learning environments, candidate agent evaluations, and automated regression gates. The talk is particularly valuable for teams running production agents at scale who need systematic improvement mechanisms beyond ad-hoc prompt edits.
📺 Source: AI Engineer · Published July 05, 2026
🏷️ Format: Deep Dive
