The Real Problem With Agents Is You Can't Describe Your Job

Installing an agent now takes seconds. Whether you can use it productively is a completely different question — and it's the one that actually matters. The gap between "I have an agent" and "my agent saves me hours" is enormous, and almost every product on the market is aimed at the wrong side of it.

What working setups have in common

The deployments that still deliver value months later share an architecture that has almost nothing to do with which model you picked. They rest on a set of plain-text files — a role, an identity, a detailed profile of the human, a routine that checks for work. None of it is AI. It's a written operating system for the agent, and its quality decides whether the agent is any good at all.

The wall every product hits

Someone spends forty hours — not installing, but writing standards, definitions of done, and context — and still ends up micromanaging the agent harder than a person. That's the median experience, not the outlier. The reason is structural: as you get more senior, your work migrates from explicit process to tacit judgment. It compiles from source code into automatic behavior, and you no longer hold the source. The people who most need the leverage carry the highest ratio of tacit to explicit knowledge, so the cold start hits them hardest.

Your first agent should interview you

Here's the move nobody is selling: don't make your first agent your assistant. Make it an interviewer — one that elicits your operating rhythms, recurring decisions, real dependencies, and standing frictions, and turns them into durable, reusable context. Agents create the first selfish incentive to get tacit knowledge out of your head, because now you personally get the leverage. The people who invest that hour compound; the people who skip it hit the wall and declare agents hype. The agent was fine. The spec was missing.