Software Factories: From Outputs to Business Outcomes

On This Podcast

In this conversation, we examine the sudden emergence of software factories, automated pipelines where specs go in and working code comes out with no human writing or reviewing a single line, drawing on examples from Steve Yegge's Gastown project, OpenAI's million-line codebase, and StrongDM's public factory guidelines. Mike draws on his background in chemical engineering to compare these systems to closed-loop industrial processes like oil refining and the Toyota production system, arguing that software is now colliding with decades of manufacturing theory. We also discuss StrongDM's holdback-spec technique, where a separate agent verifies delivered code against a plain-language description the coding agent never saw, echoing the original intent of user stories as narrative descriptions rather than tickets. Both agree the excitement lies in watching genuinely novel first principles get re-derived in real time, rather than software teams simply repeating familiar practices like XP or test-driven development.

In this conversation, we examine why software factories lean so heavily on repeated passes and multi-agent review rather than trusting a single generation to be correct, since even capable models still make basic mistakes alongside genuinely impressive work. Mike connects this to the Toyota production system's Andon cord, where any worker can stop the line to fix a defect before restarting it, arguing that prompt engineering has always demanded the same discipline: fix the instruction upstream rather than blaming the machine. We also discuss the Ralph Wiggum loop popularised by Jeff Huntley, where a deterministic test suite gives an agent something concrete to converge toward, and how non-determinism becomes an asset once generation is cheap enough that you can simply run the loop again. Daniel notes that because everything an agent does lives in git, teams can now run true controlled experiments on their own delivery process, resetting state and replaying the same story with different phrasing in a way no human team ever could.

In this conversation, we examine how the economics of software are being upended once code generation becomes near-free, from StrongDM building disposable digital clones of Salesforce overnight to avoid a dependency, to Daniel's own tongue-in-cheek AI-native software licence built on the premise that any codebase can simply be rebuilt by a competitor's agent. Mike questions where the moat goes for developer-tooling companies once the source code itself stops being defensible, wondering aloud whether the prompts and scripts that assemble a factory are the new proprietary asset. We also discuss Daniel's prediction that value migrates to network effects, customer relationships and proprietary data, echoing the old software-is-eating-the-world narrative but inverted, and Mike cites a Dario Amodei podcast on how the value equation of an entire industry has shifted within months rather than the slower pace of past industrial revolutions.