Descriptions:
At the AI Engineer conference, Armin Ronacher — creator of the Flask Python framework and co-founder of Earendil — and Christina Poncela Cubeiro present a practitioner’s analysis of what goes wrong when engineering teams over-rely on AI coding agents. The talk opens with a concrete example: a company whose tagline was “ship without friction” suffered a production incident caused by an accidentally deployed configuration change, used as a case for deliberately reintroducing friction into the development process.
Ronacher and Cubeiro identify two compounding failure patterns. The first is psychological: AI coding tools are addictive, they manufacture an illusion of productivity through high output volume, and the industry baseline has shifted so that engineers feel constant pressure to use them — leaving no time for genuine review. The second is architectural: agents optimize for making progress over correctness, producing code that silently recovers from local failures, duplicates logic already present elsewhere in the codebase, and generates entropy that eventually exceeds the agent’s own ability to navigate the project.
A key practical insight is the library-versus-product distinction: AI agents perform significantly better on library codebases, which have tight constraints, well-defined APIs, and simple cores. Product codebases — with interacting concerns like billing, permissions, feature flags, and UI — are far more prone to agent-generated brittleness. Ronacher draws on 12 months of building with and on agents at Earendil, making this one of the more grounded and technically specific talks on the limits of agentic development currently available.
📺 Source: AI Engineer · Published April 18, 2026
🏷️ Format: Keynote Launch
