What a fully autonomous multi-agent build taught us about engineering AI systems

An AI agent that builds software will not stop on its own. That is the finding underneath every guardrail described in this paper, and it came from watching it happen 11 separate times.

Led by our Director of Engineering, First Factory's deep-agents-factory ran its first complete end-to-end mission on July 9th and 10th of 2026. The goal was to build a key but discrete feature—not an entire application—that was fully defined, built, and tested autonomously by AI agents. This is the kind of work referred to as a software factory. A one-line goal went in. A working, tested product feature came out, without a line of application code written by an engineer.