Lessons from AI
First Factory recently tested the software factory model on a real feature. We gave a team of AI agents a one-line goal, a single-user Kanban board with drag-and-drop cards, and let them write the requirements, build it, test it and ship it without an engineer writing any code. They delivered a working feature, but a human still had to step in 11 times to get it there.
In his latest Inc. column, "1 Feature, 11 Human Interventions," First Factory COO Don Gregori explains what went wrong and what we built so it won't happen again. One agent repeated the same four-minute cycle 78 times. Two of the five bugs our checker agent caught had already passed the builder agent's own tests. His conclusion is that AI build speed is real, and so is the risk it scales, which is why senior engineers need to stay in the loop.
First Factory runs experiments like this because our engineers build with AI on real client codebases every day. That hands-on work is part of why Anthropic admitted First Factory to the Claude Partner Network as a Select partner.
.avif)
