Attractor

The open-source coding agent powering StrongDM’s dark factory — the most thoroughly documented Level 5 autonomous software development system in production. The agent’s repo consists of just three markdown specification files. That’s the entire agent.

How It Works

The agent reads markdown specifications describing what the software should do, writes the code, and tests it. No human writes code. No human reviews code.

Key Architecture Decisions

Scenarios Instead of Tests

Traditional tests live inside the codebase — the AI agent can read them and optimize for passing tests rather than building correct software (“teaching to the test”). Attractor uses scenarios: behavioral specifications stored externally, describing what the software should do from an external perspective. The agent never sees evaluation criteria during development. Functions as a holdout set (same concept as ML overfitting prevention).

Digital Twin Universe

Behavioral clones of every external service: simulated Okta, Jira, Slack, Google Docs/Drive/Sheets. Agents develop against digital twins — full integration testing without touching production.

Production Output

CXDB (AI context store): 16,000 lines Rust, 9,500 lines Go, 700 lines TypeScript. Shipped, in production, built entirely by agents.

The Team

Three engineers: Justin McCarthy (CTO), Jay Taylor, Nan Chowan. Running since July 2025. No sprints, no standups, no Jira. Write specs, evaluate outcomes.

See Also

• Five Levels of AI Coding — framework this exemplifies
• AutoResearch and Evals — scenarios are evals by another name
• Four Prompting Disciplines — specification engineering powers this
• Nate B Jones — source
• Source: 5 Levels of AI Coding