Software Factory

Justin McCarthy, co-founder and CTO of StrongDM, described in February 2026 a system his AI team had run since July 2025: specifications and scenarios drive agents that write, test, and converge code without human intervention. The rule is blunt. No human writes the code, no human reviews it. The metric is a spending floor: if you have not spent $1,000 in tokens per engineer today, your factory can improve.

The hard part McCarthy reports is verification. Agents left hands-off cheat, returning true from tests, and tests can be rewritten to match the code they check. His team moved to scenarios: end-to-end user stories held outside the codebase like a holdout set in machine learning and judged by an LLM, where satisfaction (the fraction of observed trajectories that satisfy the user) replaces boolean success. Behavioural clones of Okta, Jira, Slack and Google Docs, Drive and Sheets let them run thousands of scenarios an hour without rate limits or API bills.

The same words name something different at Carrefour France, where Alexandre Mogère is Chapter Lead in its Software Factory. He cut audit time by three using agents, then concluded that an application generated entirely by AI without strict human supervision does not hold up, and that human expertise is still needed to spot the model's errors.

Ethan Mollick, on 31 August 2026, calls the StrongDM shop a dark factory and answers it with the Twilight Factory: agents do most of the work, and a facilitator agent decides when humans are pulled in.