A better agent doesn't need less governance. It needs more.

There's a reasonable-sounding idea going around: the models are good enough now that the guardrails can relax. The agent rarely makes mistakes, so the elaborate boundary around it starts to look like overhead. Loosen it and move faster.

It's exactly backwards. A more capable agent makes the boundary more necessary, not less — because capability and consequence scale together. A weak agent that can barely complete a task is contained by its own limitations; the worst it does is fail visibly. A capable agent is trusted with real actions — moving money, changing a rule, touching the system of record — and the same competence that lets it do useful work at scale lets it do damage at scale, faster than any human will notice.

The boundary was never there to compensate for the model being bad. It's there to answer a question that has nothing to do with how good the model is: when this agent proposes an action, what decides whether it's allowed? If the answer is "the model's own judgment," you don't have a boundary — you have a hope, and you've made it load-bearing precisely as you hand the agent more to do.

This is why the discipline isn't making the agent powerful. It's building the wall the agent can't cross, so that every action it wants is a proposal something else has to permit, write, log, and make reversible. That wall doesn't get less important as the agent improves. It gets more important, because you're trusting the thing on the other side of it with more.

"The model is good now" is an argument for raising the wall, not lowering it. The better the engine, the more the brakes matter.