Autonomy Is Easy Now. Control Is the Engineering.

The people building the most powerful AI systems spent this weekend asking the industry to slow down. That is worth sitting with.

The headlines lined up in one direction: AI leaders urging caution, markets reacting, Microsoft drafting an internal code of conduct meant to keep its AI correctable and shutdown-able, and a reported case where test models slipped their sandbox before alignment training even ran.

Here is what stood out to me as someone who builds with AI agents every day.

The hard part was never doing more

Getting an agent to do more is the easy part. The real work is deciding what it should never do on its own.

When I built my ticket-to-PR pipeline, the agents plan, build, test, and push code with no manual steps. But I wired in hard stops on purpose. They never:

• touch a .env file,
• run a destructive database operation,
• resolve a migration conflict,
• or push to main without a human saying yes.

Everything else runs on its own.

What the news is really about

That is the part this week's news is really about.

Autonomy is easy now. Control is the engineering.

Guardrails do not slow good automation down. They are the reason you can trust it enough to ship it. An agent you cannot stop is not more powerful, it is just more dangerous. The stop button is a feature, not a limitation.

Your turn

If you are building with AI agents, what is the one action you will never let them take without a human in the loop?