Learn-Driven Development

For decades, software development optimized one thing: building the right thing in as little time as possible. Agile worked because the dominant constraint was time.

Agents dissolved that constraint. A complete feature can be built in about an hour, and rebuilt in another hour if it turns out wrong. One person with agents can run a hundred experiments in a week. But we discovered that users did not get faster. When our production pace outpaced their absorption, so many changes reached users at once that they could not keep up, trust in the product eroded, and the feedback loop lost grip. The more we shipped, the less we learned.

We tried the obvious answer, "build incrementally", but when working with AI agents, we found it has its own trap. Thin slices (smaller than a feature level) work when humans carry context between them. Agents do not. Each slice forces the agent to rebuild context from scratch, and every rebuild introduces drift that is invisible until you step back and look at the whole.

What worked for us was separating two rhythms that used to be one. Build each feature whole, in a single coherent pass where the agent has full context. Then reveal it to users at your discretion through flags, progressive rollout, canary releases. The build increment and the feedback increment become decoupled. The code itself becomes the documentation the agent reads on the next loop. That is what we call Learn-Driven Development.

Validated learning, not shipped features, is the unit of progress. The structure of the work has to respect both user absorption and agent coherence.