There is one idea from the conversation with the makers of Claude Code that, in their words, almost everyone gets wrong. It is about verification. Ask a developer what that means and you hear: unit tests, a linter, a type check. Fair enough, because those are exactly the things that are easy to automate. They were automated long ago too.
But that is not what they mean. Verification for an agent is something else, and simpler to say than to do: can the agent run the thing itself?
The model that tested itself
They tell a moment I can easily picture. With Opus 4 they hooked the model up and gave it the task: build a feature, then test yourself in the terminal. It opened its own Claude session and tested its own work. Back then that was astonishing. Now it is ordinary: loops running against an iOS simulator, an Android simulator, a real desktop through computer use.
One of the makers works on the desktop app and describes what it looks like. There is a skill that teaches Claude to start the local app. Claude opens the app, clicks through a new piece of interface to try it out, tests the edge cases, and when something is broken it fixes it and checks again. Not the human verifying that it works. The agent doing it itself.
The mistake you do not make twice
Here is the principle underneath it that stuck with me most. When Claude makes a mistake, the maker does not say: do it differently next time. He has it record the mistake, in the instruction file or in a new skill. As a result the agent does not make that same mistake again. And that is exactly why an agent can keep working almost endlessly: every miss makes the system a little better, instead of you having to step in each time.
How I do it myself
On this site my verification is smaller, but the idea is the same. Before an article goes live, a build runs, a check on the publishing schedule runs, and a local preview runs. Not because the text compiles, but because I want to see that it holds up before it goes out into the world.
A small example from making this very series. The first cover image came out photographic, while the style of this site leans on screen print. I only caught it by putting it next to an existing image and running it. Verification is nothing more than that: not assuming it is right, but making it show you.
That is the step that separates an agent you have to watch from an agent you can trust. And trust is exactly what you need to let it run.
Part 3 of a series on how the makers of Claude Code work with it themselves. Earlier: tell the model less and I dropped Plan mode.