A software factory is an autonomous production system in which agents handle the code lifecycle from requirement decomposition through implementation, testing, and deployment. People design, operate, and maintain the factory rather than writing every change, reviewing every diff, or supervising each terminal session. section 5.7
The appealing shortcut is to describe a feature and return to merged code. The practical reality remains less mature: most teams use agentic IDEs or terminal agents, while headless pipelines and multi-agent orchestration introduce approval, coordination, and safety risks. section 5.7
Operating a software factory shifts engineering leverage toward the system around code:
- Define precise task descriptions and targets instead of manually implementing every task. section 10.8.2
- Design verification that keeps generated changes honest, including tests, output review, and edge-case handling. section 5.7 section 10.8.2
- Treat orchestration, permissions, thresholds, and infrastructure as first-class engineering work rather than assuming agents coordinate safely by themselves. section 5.7 section 10.8.2
The dark-factory tradeoff
StrongDM's three-person AI team proposed that humans should neither write nor review code, with a target of at least $1,000 in daily token spending per human engineer. The logic treats agent generation and validation as cheaper than human review, but the economics require a product able to carry roughly $20,000 per engineer per month in inference costs. section 9.1.3
