When your AI moves without you: the check that catches it

Heads up: this one's an ad.

Your documentation is a product decision. The companies building the most-used developer tools in the world chose Mintlify because great docs reduce support load, accelerate adoption, and convert evaluators into customers. Backed by a16z and Salesforce Ventures, Mintlify powers 20,000+ companies with AI-native documentation that keeps pace with your product — without pulling engineers off the roadmap to maintain it.

Book a demo

The one that got me was a Friday morning, running late for a call. I pasted a task into an agent without stopping to set the scope. No ask before you move. No flag for my review. It ran 3 steps before I caught up.

The output was usable. What I'd skipped were the explicit boundaries that tell it when to stop and wait. That was the lesson. I've found it happens in 2 ways: when I'm rushing and miss the guardrails entirely, and when I'm sure I've done everything right and something still slips through. Now I set them every time. If I don't want it to act without my sign-off, I say so. If I want to see everything before it goes anywhere, I say so.

Last week I wrote about steering AI output vs. letting it run. A few of you came back with the follow-up: what happens when the agent doesn't wait to be asked? That's what this issue is about.

I gave an agent one standing instruction: monitor my calendar for meetings without a prep brief, draft one from our past notes and research, and then flag it for my review.

What almost stopped me was that I wasn't sure I could trust it to get the context right without me guiding every step. It takes time to dial in how well an agent performs. They make mistakes the same way any workflow does when you're still calibrating it. The difference is learning to build your review process around that reality instead of waiting for it to surprise you.

It ran 3 times before I checked in. 2 of the briefs were solid. One was confident and wrong. It had pulled context from the wrong notes entirely, and it had no idea.

That brief is the whole job. The agent did the work. I still had to know enough about the situation to catch that it had done it wrong. An agent that acts is only as reliable as the person reviewing it.

Once again, it’s the judgment layer.

The prompt is for CoS Signal members, along with the upcoming prompt library and everything else members get. Founding rate is locked at $14.99/month or $99/year.

Gartner predicts 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. The gap between those 2 numbers is the space where governance decisions happen. Chiefs of Staff live in that space.

Anthropic published research on how people actually use agents in practice, specifically where agents ask for help and where they move on their own. The key finding is that as experience grows, effective oversight shifts from approving every step to intervening at the right moments. Worth reading before you set up anything that runs without you in the loop.

We're looking for guest contributors! Have you figured out how to use AI to do your job better, and you've been meaning to write it down? This is the nudge.

What we're looking for: specific workflows, real problems solved, honest takes on what broke before you got it right. The more concrete, the better.

What we offer: editing support, your byline, and distribution to our subscriber base.

Heads up: this one's an ad.

Most e-commerce sellers juggle 6–8 tools and pay hundreds monthly to keep operations running. StoreClaw replaces the stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks profit 24/7. Connect your store and let AI handle the work — no prompts, no complex setup, no credit card required.

Try if for free today