Count Your Agent Mechanics

If I could ask a large company one question to judge its AI readiness, I would skip the model vendor and the pilot count. I would ask how many agent mechanics it has. I mean people whose full-time job is building, tuning and repairing agents. At a lot of thousand-person companies the answer is zero, and with zero, the company can only run its legacy business with better autocomplete.

This is the metric because agents break in boring, constant ways. A network of agents doing a hundred jobs will do some of them wrong every day. Instructions drift away from the process they describe, a token expires, an upstream system changes shape. Someone has to be turning screws. Model quality is mostly good enough for everyday work. Staffing is where companies are stuck.

It also matters because building agents is software development, whether or not anyone calls it that. Ask an assistant to analyze a spreadsheet and it writes Python on the spot.^[1] Wire a few steps into a workflow and you now own code that needs review, testing and a place to run.^[2] I have argued that automating knowledge work is software engineering, and the mechanic is the person who carries that discipline into teams that never thought of themselves as shipping software. It is the same reason to budget for agent owners and not just the builds.

So make the goal concrete. Pick ten people this year and move them from whatever they did before into full-time agent building. Give them a cloud environment where agents run with scoped access instead of on someone's laptop, a review process for what the agents produce, and a real mandate. Ten is small enough to start next quarter and large enough to change what the company believes it can do.

Then track the number. Count how many of those ten are running agents the business relies on by year end, and how many more people have asked to join them. That count will tell you more about your AI readiness than your spend report.