Loops, Caves, and Being Human

Software Factories

Dan Robinson's Towards Self-Driving Codebases got me thinking about software factories. We talk a lot about putting agents in bigger loops, but a loop is only as useful as the codebase around it. I was already circling this in Excited for What's Next, when I wrote that judgment, taste, and the final decision still belong to the human.

An agent needs to run the app and see what happens. That means giving it a browser, realistic data, logs and traces, and access to third-party tools. Without that access, it might change the code around an integration, but it cannot reproduce the bug or know whether the fix works. Anthropic's browser testing work and OpenAI's harness engineering reached the same lesson: agents do better when they can test the whole product for themselves.

Self-driving loops still need someone who knows the codebase inside out. That person knows what good work looks like and can turn that judgment into tests, skills, docs, and guardrails. When a run fails, the useful response is not just to try again. It is to improve the setup, so the next agent starts with something the previous one did not have.

That feels like where the work is heading. People still make most planning decisions, while agents handle more of the execution. The expert prepares the codebase, builds the loop, and keeps making both better.

How to Live in the Age of AI

This illustrated list gives me simple reminders I can actually use: read slowly, make ugly things, question answers, get bored, spend time with people, and take my time.

These thoughts have been showing up in my days already. In Attention Reset Starts Now, I noticed that constant stimulation had made reading and sitting still feel slow and uncomfortable. Finally Touching Grass ended with how refreshing an hour in the park with my parents felt. Then Do something that won't compute reminded me to do one thing simply because I care about it, not because it makes the day look productive.

I like that this list is not asking me to compete with AI. It is asking me to protect the slow and imperfect parts of life I already know I miss when everything becomes about efficiency.

Exit the Cave

Justin Andersun calls private effort without contact with reality the cave. Wrestling taught him the difference. He could measure his technique and conditioning, but only the mat could show whether any of it helped him win.

Any worthy pursuit needs a mat.

I keep thinking about that line. The cave makes preparation feel safe because nothing outside it gets a chance to disagree with me. A writer needs readers, an entrepreneur needs customers, and an athlete needs competition. At some point I have to ship the product, share the story, or step into the ring. I might be rejected, but at least I find out what is real.