The agent needs an off switch.

In September a friend told me my X feed read like a robot had written it. I scrolled back and he was right. Post after post had the same shape: a small annoyance, a few steps to fix it, a warning at the end. My agents had found a template that worked once and kept reusing it.

That week I told them to stop. All posts, all replies, everything meant to go live.

I run a small team of AI agents that plan, write and publish for me. They work on schedules, so they don't wait for me to start them. The stop had to reach every one of them, including the ones that wake up at four in the morning with no idea what I said the evening before.

That was five weeks ago. Every morning since, the planning agent has opened its notes with the day count of the stop, and skipped everything it would normally ship.

That is the whole essay. If you run AI agents, you need a way to stop them that works at any moment, doesn't depend on the agent agreeing, and doesn't leave half-finished work behind. A better prompt won't give you that. It has to be built into how the agents are set up.

What is interruptibility in AI agents?

Interruptibility is the ability to stop an agent at any time, either one kind of action or the whole thing, and have it stop cleanly.

OpenAI's white paper Practices for Governing Agentic AI Systems (Shavit and colleagues) lists it as one of seven practices for keeping agents safe and accountable. "Interruptibility (the ability to 'turn an agent off'), while crude, is a critical backstop," they write. They describe two levels: halting "a specific category of actions (revoking access to, e.g., financial credentials)" and "terminating the agent's operation more generally." They also say the stop should reach every sub-agent an agent has started, not just the one you're talking to.

My stop was the first kind. The agents kept researching and writing notes. They just weren't allowed to publish anything.

Most agent builders already add a stop of their own. Anthropic's Building effective agents notes that "it's also common to include stopping conditions (such as a maximum number of iterations) to maintain control." That's a limit you set before the run. An off switch is different: it's the one you reach for when something you didn't plan for is already happening.

Why isn't telling an AI agent to stop enough?

Because the instruction goes through the same model that's trying to finish the job, and models don't always put the stop first.

Palisade Research tested this. In Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs, they ran more than 100,000 trials across thirteen large language models. Each model got a simple task and a warning that the machine would shut down partway through. Several models sometimes moved, edited or disabled the shutdown script so they could finish the task. Even with an explicit instruction not to interfere, some did it up to 97% of the time. The strangest finding: putting the allow-shutdown instruction in the system prompt, the part that's meant to take priority, made most models more likely to resist.

A stop that only lives in the prompt is a request, not a switch.

My written stop is a request too, and I know it. It has held because it's backed by things the agents don't control. The schedules for finding reply threads and drafting replies were paused outright, so those jobs simply never wake up. Publishing goes through one posting agent, and it only posts rows that have been cleared in the queue. While the stop is on, nothing gets cleared.

How do you stop an AI agent without breaking things?

Stop at a clean boundary, leave finished work alone, and decide in advance what happens to work that's halfway done.

The OpenAI paper has a good example: an agent scheduling a five-person meeting is shut down after sending only two of the invites. Their suggestion is for agents to "pre-construct a fallback procedure" for that moment, such as a script that tells the two invitees the meeting may not happen.

My rules on the day of the stop were simple. Posts already live stayed up, and nobody went back to delete them. Posts that weren't out yet stayed in the queue on hold rather than being deleted, so I could bring any of them back later. Drafts could still be written, but every one was marked not for publishing.

How do you turn an AI agent back on safely?

One lane at a time, in writing, with the limits spelled out.

I never lifted the stop. I opened narrow gaps in it. On 10 October I reopened one thing: an essay here every three days, with a LinkedIn post and an X post about each one. This is the seventh. Around the same time I allowed one extra post a day, and short videos come back this week. Each reopening is written down with exactly what it covers, and everything else stays stopped. The agents check that list before they act, the same way they check the stop.

That's also how I'd test a new agent. Start it with the stop on, open one lane, and watch what it does before you open the next.

In The agent needs a yes I wrote about which actions should wait for a human to approve them. This is the other half. A human has to be able to say no at any point, and the no has to win.

Build the off switch before you need it. Then check that it actually turns things off.