Self-Driving Train
2026-05-28
Trains amirite?
Anyways, the world is abuzz discussing approaches for teaching, guiding, even scolding these models to produce higher quality output.
Rules files, prompt engineering, dumping more files and data into their context. Setting up evals.
It's fun / people feel productive, companies have content to build around it. And no surprise if you tear this all down to the studs, and simplify the naming we are back to SDLC best practices. Funnily now everyone is kind of becoming an expert in that dreaded TDD ? Write the spec first, design tests for the feature/code you are going to have the model write. Keep your documentation "colocated" with the code, and keep it up to date.
This time/work spent hoping to make it impossible for the models to mess up, reminds me of a few things, one being falling into the pit of success:
And talking with colleagues about a metaphorical self driving car for how agents/models are working these days. Tell it where to drive then sit back. We then moved onto the self driving train. What I want is to get from A to B fast, safely and predictably. Maybe you don't want that and you want more of a road trip with many rest stops if you have lots of time and no deadlines. And go ahead, by all means it's fun designing your own highways and all that.
But if we want to move faster with predictable outcomes then we really just impose the same constraints & best practices we have been applying to engineering forever. Leaning into typed languages, typed api contracts, linting, automated e2e testing we can get closer to this self driving train. And the models will better "one shot" your desired outcomes.
Developers love to get side tracked building themselves a car when the tracks are already there ready to rip.