Why AI Agents Need Room to Fail Before They Learn

Giving an AI agent real autonomy means accepting it will fail early and often before it gets good, the same curve organizations hit during any real change. Peter Maddison brings a stuck OAuth problem to the table: an AI agent that kept going in circles and couldn't find its way through. That leads into a conversation from Dave Sharrock's local AI meetup about an AlphaGo-style approach to AI agent autonomy: instead of specifying every step, you define hard constraints and let the model work out its own strategy inside them. Peter and Dave connect this to the Virginia Satir change curve, the same dip in performance that shows up when an organization tries a new way of working, and to the difference between using AI to optimize what you already do versus using it to rethink the business itself. They also get into how experiments like Andon Labs' AI-run cafes and vending machines use small dollar constraints to let a model learn from failure without real financial risk. This week's takeaways: \- A well-articulated objective with clear guardrails lets an AI agent find its own path to a solution, even one you didn't expect or fully understand. \- Real learning, whether it's an AI agent or an organization adopting a new way of working, comes with an unavoidable dip in performance that can't be planned away. \- The bigger opportunity with AI isn't squeezing more efficiency out of an existing process, it's using AI to test entirely different ways a business could operate. Listen to the full episode at definitelymaybeagile.com Subscribe so you never miss an episode. Have a question or topic you'd like us to cover? Reach out at feedback@definitelymaybeagile.com