What Matt Pocock gets right about coding agents

Matt Pocock tried to run his favorite project from detailed specs. For months. The agent's code got worse every week, and he kept going anyway, because the idea came from an OpenAI engineer and surely he was the one doing it wrong.

Then he quit the specs, went back to the code, and things started working. On Nate Herk's podcast he explains why: the agent was never the weak link. What you leave unsaid is, and so is the place the agent works in. Here are four lessons from the conversation, and where each one lands in The Agent Office.

1. Your agent doesn't know what you care about

Pocock says people underestimate the gap between what you mean and what the agent knows, and that no model is smart enough to close it. The dangerous part is what you don't know to say: security, how it gets deployed, the thing nobody thought of yet. Leave those out and the agent makes something up, or decides they can't matter because you never mentioned them.

His fix is a skill called grill-me. Before any code, the agent interviews you until the blind spots are out of your head.

In The Agent Office: that interview is the Planner on the Idea board. You talk the idea through, it drafts the ticket, and then you hire a worker. A missed assumption costs two minutes of chat, not a branch.

2. The codebase is the agent's environment

This is why the specs failed. Pocock's point is that code is where the agent lives, and an agent in a rubbish codebase produces rubbish. His test: how would you lift the weakest developer in a team of a hundred? Good tests, checks that stop whole classes of bugs, a codebase where a change won't break what nobody thought about. He also made his app runnable by the agent, so it could click through it and take screenshots. He says the jump in quality was large, and that with that setup an older model would have done just as well.

In The Agent Office: every worker starts from nothing, so the office hands it the lessons. It searches your repo's saved lessons and puts the relevant ones in the brief, and a worker saves a new one whenever you correct it. Skills you mark "Always for workers" go into every brief, and so does a self-review checklist. Fix a mistake once and the next worker doesn't make it.

3. Raise the floor, then turn the effort down

Pocock runs Opus 5.5 on medium effort and is thinking about low. Most people, he says, crank effort up on the assumption that more is better. His agents don't need to work that hard, because the tests catch what they get wrong, and the answers come back faster and cheaper.

In The Agent Office: each worker has Model and Effort dropdowns. Run the small fixes light and save the heavy settings for the hard tickets.

4. Climb the trust ladder, brake first

You shouldn't trust an agent the first time it works, he says. When you learn to drive, you learn the brake first. So watch the agent like a hawk, and only then add a second one, or a few skills. And start it on the boring work you dislike, where you can see right away whether the output is good.

In The Agent Office: the brake is the default. A worker asks before it pushes or installs anything, and you approve with a click. On risky changes (a big diff, sensitive paths, changed tests) the office hires a separate reviewer on a different runner, which comes back with approved, changes requested, or needs you. Start with one boring ticket and one worker.

Ship more, without shipping worse

Every builder wants to be more productive. Pocock's answer is that the lever is not a smarter model or a longer prompt. It is four habits: say what you care about, give the agent an environment that catches its mistakes, spend effort only where the work needs it, and earn trust one rung at a time.

Those habits are easy to agree with and easy to drop when you are busy. The Agent Office is where they stay in place. The Planner interviews you before work starts, lessons carry from one worker to the next, effort is set per worker, and approvals and a reviewer sit between every agent and your main branch. You stay the manager, and the office keeps the habits running while you build.