Software Factory paradigm: no-human-review coding with scenario-based validation¶
Insight: StrongDM's AI team operates a "Software Factory" where code is neither written nor reviewed by humans. Specs and scenarios drive coding agents, with validation through "satisfaction" metrics (probabilistic success rates across observed trajectories) rather than binary test passes. Their benchmark: "$1,000/day per engineer in token costs" indicates the economic threshold for this approach.
Detail: Key innovations include: (1) Scenario testing as "holdout sets" stored outside the codebase so agents can't cheat; (2) Digital Twin Universe — behavioral clones of third-party services (Okta, Jira, Slack) built by agents for unlimited testing without rate limits; (3) Attractor, their coding agent released as spec-only markdown (no code); (4) Novel terminology: "Gene Transfusion" (extracting patterns between systems), "Semports" (cross-language porting), "Pyramid Summaries" (multi-level summaries for agent context). Founded July 2025, the team identified Claude Sonnet 3.5 (Oct 2024) as the inflection point where "long-horizon agentic coding workflows began to compound correctness rather than error."
Sources
Related: existing entry "Agentic coding workflow maturity levels" in external/claude-code.md — COMPLEMENTS