“Software factories” and “Harness Engineering” are two names being applied to parts, or the whole, of a system of AI agents that produces software as its output. Since these techniques could be used to produce something else, they’re a reasonable model for understanding what AI agents are capable of, and the work involved in making them produce something useful.
StrongDM is a security software company, and they have a team that approached this with some principles (e.g. “No hand-written code”) that I’ve heard from other pioneers. Read Software Factories And The Agentic Moment.. Then check out Attractor, their implementation of coding agent.
Ryan Lopopolo at OpenAI describes a very similar approach to StrongDM’s in Harness Engineering: Leveraging Codex in an agent-first world
Mitchell Hashimoto is the author of Ghostty, a very nice cross-platform terminal emulator. His blog post, My AI Adoption Journey should resonate with a lot of us, in whole or in part.
Birgitta Böckeler comments on several of the above blog posts in Harness Engineering, and she zeroes in on the importance of constraints.
For a critical view, read Tom Wojcik’s What AI coding costs you.
Finally, Simon Willison is collecting best practices in Agentic Engineering Patterns.