Stop with the Agents!

Why custom, declarative software will eat the world. Not a chain of non-deterministic agent bots.

So I'm doing the LinkedIn thing and seeing way too much of the word "agentic" or "AI agent", while watching mostly hype with few real implementations that wow me. Meanwhile, actual developers are just talking about how Cursor or Claude Code got their side project running in a couple of hours.

When the devs are talking about something else, that divergence is a signal.

When I looked at LangChain or CrewAI, I started getting early game dev feelings. I don’t think people quite remember what scripting NPCs in a game engine is like.

LLMs are inherently non-deterministic. Sometimes they do something a little different each time... by design. If you’re chaining these things together, you need to add a lot of glue to make it work. It’s not easy. And the UIs I’m seeing for chaining? They give me NPC bot scripting flashbacks. (Look Ma! Flows not code! <shudder/>) And Dear Lord, I do not want to be the one debugging those chains.

Meanwhile, Cursor generates me a script that calls out to an LLM at specific points. I can seal those off, put checks around them, and end up with something far more deterministic and simple.

But what if things change? Isn’t that where agents shine?

Maybe. But here’s the first thing I’d try instead:
- Save your specs.
- Update your specs.
- Have Claude regenerate the script.
- Test. Deploy. Observe.

We have a hammer problem with agents. And now everything looks like a nail.

All of this leads to something else. The age of bespoke, declarative software is coming. And as a developer, I’m happy.