StrongDM is known here through its AI team, which published the first public description of the working arrangement they call a Software Factory: non-interactive development in which specifications and scenarios drive agents, under the team’s stated rules that code must not be written by humans and must not be reviewed by humans. The software this team was building manages user permissions across a suite of connected services, which How StrongDM’s AI team build serious software without even looking at the code singles out as notable: security software is the last thing one would expect to be built from unreviewed agent-written code.
The team is a distinct and recent part of the company rather than the whole of it. It was founded in July 2025 on the rule of no hand-coded software, and consisted of three people at the time it was first seen working. Its own account of why it formed when it did points at a transition observed in late 2024: with the second revision of Claude 3.5, the team state, long-horizon agentic coding workflows began to compound correctness rather than error, and by December 2024 the model’s long-horizon coding performance was unmistakable to them.
History
What is recorded here is the AI team’s history rather than the company’s. It was founded in July 2025 and had been running three months when it demonstrated its work to a small group of invited guests in October 2025 — by which point, on that visitor’s account, it already had a working coding agent harness, clones of half a dozen third-party services, and a swarm of simulated test agents running through scenarios. That was before the model releases at the end of 2025 that the same account credits with making agentic coding significantly more reliable. The public write-up followed in early February 2026.
Activities & Products
The team’s distinctive engineering output is the apparatus that makes unreviewed agent-written code tolerable to them. The Digital Twin Universe is their set of agent-built behavioural clones of the third-party services their software depends on — twins were built for Okta, Jira, Slack, Google Docs, Google Drive and Google Sheets — which they use to run thousands of scenarios per hour without rate limits or API costs. Alongside it they use scenarios held outside the codebase as a holdout set, and a probabilistic measure they call satisfaction in place of a green test suite.
Two pieces of software were released with the write-up. Attractor is the non-interactive coding agent at the heart of the arrangement, published in an unconventional form: the repository contains no code, only markdown files specifying the software and a note telling the reader to feed those specifications into their own coding agent. cxdb is a conventional release — an AI context store holding conversation histories and tool outputs in an immutable DAG. The team also publish a set of named techniques, among them gene transfusion, semports and pyramid summaries.