The most powerful tools we ever built, and never learned to use

Unix gave us sed, awk, grep, find, xargs and later git. Between them they can restructure a million-line export, reconcile two data sets, or reconstruct the history of a change in a single line of text.

Almost nobody ever used them that way.

Not because the tools were weak. Because the syntax was a wall. Learning to express a real problem as a regular expression took years, the knowledge was not transferable, and every engineer I know — myself included — settled for maybe five percent of what those commands could actually do. The rest sat there, fully implemented, fully documented, unused for fifty years.

That wall has just come down.

An AI agent with a shell can write the invocation. It knows the flags, the escaping, the pipeline order, the difference between GNU and BSD behaviour. You describe the outcome; the tool does what it has always been able to do. The interesting part is not that this saves typing. It is what kind of automation it produces.

Three things matter here, and they matter especially in industrial and OT work:

Recommended by LinkedIn

The tools are deterministic. sed has no opinion. The same input and the same expression give the same output, every time. The model supplies the instruction; the tool supplies the guarantee. That is a very different risk profile from asking a model to reason about your data and trusting the answer.

The result is inspectable. Text on a filesystem, under version control, diffable and re-runnable by someone who was not in the room. An audit trail as a by-product, not an afterthought.

Fifty years of installed base comes back into range. Every SCO and HP-UX machine, every engineering station, every historian export and every log directory that was "too painful to process" is now workable. In plants where the control system is older than the engineers maintaining it, that is not a productivity gain — it is access to evidence that was previously out of reach.

This is the shape of AI I can actually defend to a utility: not an autonomous advisor forming judgements about a process, but a steered instrument driving deterministic tools over a bounded data set, with every step written down and repeatable. Verifiable, not trusted.

We spent decades assuming the next leap would come from new tools. It came from finally being able to operate the ones we already had.

If you work with legacy Unix systems, historian exports or control-system data: this is worth an afternoon of your time. The tools are already installed on your machine. They always were.