Late Stage Agentic

How not to make a mess of agentic coding.

I keep asking myself this, and I keep getting asked it: what are all you developers for, if an agent can write GTA 6 from a single prompt?

My Claude and I have been trying to answer that with something other than philosophy. There is a file that Claude writes to after every review of code it wrote — what exactly it got wrong, generalised and grouped with everything before it. The most frequent finding in it reads:

What it was handed, it treats as fixed

Solving a well-posed task reliably, predictably and elegantly is one talent, and the models get closer to it month by month — though there is a great deal still to work on, and a great deal of this site is about that. Seeing that the task is posed wrong, or that it is the wrong task altogether, is the other talent. That one, as Mastercard says, is priceless.

So inventing the skills and the methods that let an agent and a person help rather than hinder each other is the right work, and most of what gets written here is that work. But don’t count on the agents’ ever-improving analysis, planning and self-correction abilities to replace what we meat sacks get as a bonus for being limited and not lasting very long — the bumps and bruises collected over a life. And the new bruises, for every one of Claude’s mistakes, are still yours to collect.

Welcome to Late Stage Agentic — a diagnosis best heard while there is still anyone to make it.