← back
Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI
Takeaway
Turn failures into replayable learning environments before deciding which agent layer to change.
Summary
- Continual learning can improve an agent's model weights, harness, or memory, but requires both useful feedback and a method for acting on it.
- Production logs plus automated or expert feedback are insufficient without replayable environments that define tool behavior, simulated users, and success criteria.
- Fine-tuning and reinforcement learning modify models; prompt search and trace-to-harness methods modify orchestration; memory stores corrections and reusable skills.
- Unverified edits from individual traces can introduce hidden regressions, so the learning system should seek the smallest durable change and test it across scenarios.
continual-learningagent-simulationharness-engineering
Original description
Agents fail in production in ways that static benchmarks cannot fully capture. The key question is whether they can learn from those experiences without drifting or breaking prior capabilities. This talk introduces verifiable continual learning for AI agents: a framework for converting traces, failures, and feedback into testable, regression-aware improvements. I will discuss four core requirements: turning failures into replayable learning environments, preserving prior capabilities during updates, routing repairs to the right layer of the agent stack, and keeping the learning loop efficient enough to run continuously. We will use these principles to examine today’s approaches, including prompt optimizers, memory consolidation, coding-agent harness repair, and trace-to-harness systems. I will then discuss the remaining gap: a holistic, lifelong, verifiable learning loop with online regression control. Speakers: Soheil Feizi (RELAI): Dr. Soheil Feizi is the Founder and CSO of RELAI and an Associate Professor of Computer Science at the University of Maryland, College Park, whose work focuses on the reliability, safety, and optimization of AI systems. X/Twitter: https://x.com/FeiziSoheil LinkedIn: / soheil-feizi-b14a4895