World's Fair 2026
Continual Learning for AI Agents: From Failures to Durable Improvements - Soheil Feizi, RELAI
Overview
This talk introduces verifiable continual learning (VCL) for AI agents, a method to achieve durable improvements from experience without forgetting past performance. It addresses the challenges of obtaining feedback from production logs and acting upon it effectively. VCL aims to transform failures into testable scenarios, ensuring that improvements are verified and do not introduce regressions.
Who should watch
- AI Engineers
- Product Managers
- Builders working with AI agents
- Those facing challenges with agent performance degradation
- Teams looking to implement robust learning loops for AI agents
Key takeaways
- Continual learning for AI agents involves learning from experience through interaction, feedback, and improvement without forgetting.
- Two fundamental challenges in continual learning are acquiring feedback (knowing if the agent performed well and what it should have done instead) and acting upon that feedback to optimize the agent.
- Production logs are not inherently learning environments; they need to be transformed into replayable simulation and evaluation environments.
- Improvements can occur at multiple layers of an agent: the model layer (weights), the harness layer (prompts, skills, tools), and the memory layer (storing facts and skills).
- Verifiable continual learning focuses on making every fix proven to help and proven to break nothing that already worked, involving executable tests, measured deltas, and regression tests.
- Four key principles of practical VCL are replayability (turning failures into runnable tests), holisticness (routing fixes to the right agent layer), lifelongness (improving new cases without breaking past ones), and efficiency (ensuring the learning loop runs frequently and cost-effectively).
- Rely.ai offers a VCL engine that transforms signals into replayable environments, performs root cause analysis, and uses regression-aware optimization.
Notable quotes
*Humans learn mainly from experience by interacting with the world and getting feedback. The goal of continual learning is to imitate the same for agents.*
*Production logs are not learning environments. We need to transform them into replayable learning environments to simulate and evaluate the agent on the same patterns and scenarios.*
*The frontier is regression-aware continual improvement, where when fixing the new failure, we verify that we don't forget the old ones.*
Unofficial community note. Prefer the recording for nuance.