I gave an AI agent shell access. Here's what I learned.

For one week I let an AI coding agent work directly in a project repository — reading files, editing code, running builds. I wanted to find where the ceiling actually is, not where the demos say it is.

The ceiling is not intelligence. The agent reads code well, reasons about architecture decently, and writes competent implementations. The ceiling is verification. Without a fast, trustworthy way to check its own work, the agent drifts: it fixes one thing, breaks a subtle other thing, and reports success because nothing told it otherwise.

The three things that changed everything

1. a test suite that runs in seconds — slow feedback kills autonomous work
2. a small, enforced scope per session — "fix the pagination bug", never "improve the app"
3. a hard rule that git state is sacred — branches for everything, no direct pushes

What I would still never do

Let an agent touch production data, deploy on its own, or run with secrets it doesn't need. Autonomy in software should be proportional to how cheaply you can reverse a decision. Reversible? Go wild. Irreversible? Keep the human in the loop, permanently.

Agents are best understood as very fast interns with no memory of yesterday. Managed that way, they're excellent.