- Software Factories And The Agentic Moment: One company decides to run an experiment where humans don’t write any code. They can only prompt agents or set agents up to prompt eachother.
- Patterns they identified
- “Digital twin universe” in which they clone software they use to better simulate
- “Gene transfusion” where ideas in one project are copied into another
- “Shift work” where you try to move a task that needs a human into an agent loop somehow
- And the harness that makes it possible (allegedly): Seed -> Validation harness feedback (closed loop system). You should always start with a seed which can be a spec or high level description of the goal which you use to create a spec or an existing code base you can build up context with in an agent.
