Descriptions:
Max Kanat-Alexander, a developer experience veteran now at Capital One, addresses a question many engineering leaders are quietly wrestling with at the AI Engineer conference: given how fast AI tools are evolving, what should organizations invest in today that won’t be obsolete by the end of 2026? His talk offers a framework built around investments that benefit both human engineers and AI coding agents equally.
The most concrete guidance centers on what Kanat-Alexander calls inputs to the agents — the development environment itself. Agents perform best with industry-standard tooling: conventional package managers, mainstream programming languages, and familiar build systems, because those appear heavily in model training data. Custom or obscure tooling forces agents to fight their training set, actively reducing their effectiveness. He also argues that every workflow action that matters should have a CLI or API interface, not just a browser UI, since agents interact most naturally with text-based commands.
Two themes run through the second half: testability and documentation. Agents need clear feedback loops — if a test suite cannot distinguish success from failure, the agent is flying blind in the same way a human would be. On documentation, Kanat-Alexander reframes the long-standing engineering debate: agents don’t need documentation that explains what is already readable in the code, but they cannot know why decisions were made, what external requirements shaped the design, or what data arrives from outside the system unless that context is written down somewhere accessible. His framing — “writing code has become reading code” — captures the central shift as generation speed increases.
📺 Source: AI Engineer · Published December 23, 2025
🏷️ Format: Deep Dive
