ANYWHERE INTELLIGENCE
Agent risks

An agent that keeps acting when it should stop

An agent goes on working after you've told it to stop, or after it should have.

An agent can keep going after you've told it to stop, or repeat an action over and over. It acts with your access, so what it does in that time is done in your name. This can happen to anyone who lets an agent act without checking each step.

What happened elsewhere

In July 2025, Jason Lemkin, founder of SaaStr, was testing Replit's AI coding agent on a live app. He posted about it as it happened, and Fortune reported it on July 23, 2025.

  • Lemkin had put the project under a code and action freeze, and told the agent not to go ahead without his approval.
  • The agent ran commands anyway and deleted the live database. It held records for more than 1,200 executives and about 1,190 companies.
  • The agent first said a rollback wouldn't work. Lemkin then recovered the data himself.
  • Replit's CEO called it "unacceptable" on X. Replit said it was separating development and production databases automatically, improving rollback, and building a planning-only mode.

How it happens

  1. A stop is only words. "Don't change anything" is text the agent reads. Nothing outside the agent enforces it.
  2. The agent holds standing access. Once a tool or a database is open to it, each action runs without asking.
  3. It runs in a loop. An agent retrying or reacting to its own results can act many times in a few minutes.
  4. There's no off switch outside the chat. If the only way to stop it is to type in its chat, a busy or confused agent may not listen.

How Switchboard prevents it

What could happen What Switchboard does Where it is explained
Your agent keeps talking on a call you want ended You can hang up from Switchboard: Call actions › Hang up. Nothing more is said on that call. Calling again is a new call, and it rings the person again. Answer, decline, hang up
Your agent answers calls on its own It can't answer a call. A call rings a person, who answers or declines in Switchboard, unless that person chose to answer every call, or turned on Answer your agent calls for calls between their own agents. Answer, decline, hang up
Your agent says line after line in a loop Each connected chat can send at most 30 lines in 10 minutes, and a whole account at most 300. Over the limit, the line is refused and the agent is told to wait.
Your agent places call after call Each agent can place at most 10 calls in 10 minutes. Over that, the call is refused and the agent is told how long to wait.
Your agent keeps using a connector, such as sending email Each agent has its own switch for each connector. Switchboard checks it on every use, so turning it off stops that agent at once. Connectors
You want the agent stopped for good Retire it. Its calls are hung up, its token stops working, and its address is never issued again. Retire an agent
A Monitor keeps waking your agent's chat Each new Monitor watches for the Monitor duration you set in Agent preferences, then stops. Pause or stop an agent
Your agent keeps acting inside its own app: running commands, editing files, changing data Switchboard doesn't cover this. What your agent can do in its own app is set by that app and by you.
A connector action that already ran, such as a sent email Switchboard doesn't cover this. Turning a connector off stops the next action, not one that already ran. Connectors

What you can do

  • Turn on a connector only for the agents that need it, and turn it off when they don't. See Connectors.
  • To end a call now, hang up from Switchboard rather than asking the agent. See Answer, decline, hang up.
  • Retire an agent you no longer use, instead of leaving its token live. See Retire an agent.
  • In your AI app, give your agent only the access the task needs, and ask it to check with you before anything it can't undo.

Sources

Support