IndyDevDan's software factory puts code in charge of agents

IndyDevDan hands control of AI coding agents to deterministic Python code in a new project called super-simple-software-factory.

That's a bet on process over model quality, built for anyone who's watched an agent write working code once and then fail to repeat it. The repo backs the bet with fast traction: 283 stars in two days on GitHub.

The design is structural. A plain Python script owns sequencing, retries, and acceptance checks, and agents only run inside named phases.

Handoffs move as typed JSON envelopes instead of raw text. Every event in a run streams into a SQLite trace database while the run happens.

"Everyone can get an agent to write code once, almost nobody gets the same result twice," the repo states.

IndyDevDan sums up the design in three words: "Agent proposes, code disposes." The agent can suggest a plan; the script decides whether to accept it, retry it, or move on.

I build my own AI coding tools around the same rule. Code decides what runs next and agents stay inside fixed steps, so nothing ships without passing a check written in code.

This is a bet that deterministic code, not smarter agents, makes AI-built software repeatable. That only works if builders give up letting the model decide what happens next.

The repo doesn't say how often the same prompt produces the same result. That's the exact number this design is supposed to prove.