How do you stop an AI agent that's going wrong?

Why turn things off before revoking access?

Revoking a key is quick and feels decisive, but it only blocks the agent's next call. A feature it rolled out stays on, a machine it started keeps running, and an offer it sent stays open. Flagger.gg's docs make the same point about agent tokens: revoking one doesn't undo what the agent already changed, so kill first, then revoke. Work through the effects, then the access.

How do you turn off what the agent shipped?

If the agent's changes shipped behind feature flags, kill the flags. In Flagger.gg, tag an agent's writes with its run, by setting FLAGGER_RUN or passing --run, and the audit log can then be filtered by that run to show every flag it touched, with the state before and after and the agent's note. Then kill each one:

flagger kill search_v3 --note "p95 latency doubled after run_8812" --json

Flagger.gg says a kill works on protected flags, skips approvals, is exempt from per-agent rate limits, and reaches your apps within a second. Read its guide to kill switches, and feature flags instead of staging for why every agent change should land behind a flag.

How do you stop the machines it started?

Stop them, don't delete them yet. On Sandboxer, stopping a sandbox saves its memory as well as its disk and costs nothing while stopped, so you can resume it later to see exactly what the agent was doing. Delete it once you've looked; a stopped sandbox's data is deleted 14 days after it last stopped anyway, unless the sandbox was created before October 29, 2026 (those are kept until you delete them).

sbx stop <sandbox-id> --wait
sbx delete <sandbox-id> --yes

Every sandbox also stops on its own one hour after it starts unless that deadline was changed, which limits how long a forgotten machine can run. See pause and resume.

How do you stop money it set moving?

Cancel what hasn't been paid for. On Transformer, offers go out unpaid, and a brand can withdraw any offer a creator hasn't accepted yet. Accepted offers are binding, though one whose card charge failed and isn't paid within 72 hours is cancelled automatically. Offers a creator accepted and the brand paid for are held, not released: the money goes to the creator only when the Transformer team confirms the post, and after payment either side can open a dispute, which freezes the payment until an admin releases or refunds it. See pay on confirmed posts.

How do you cut off the agent's access?

• Sandboxer: sbx api-key revoke <key-id> stops a key at once. Its sandboxes keep running, which is why you stop them first. An AI app connected over MCP is listed under API keys, and Disconnect ends its access at once.
• Flagger.gg: revoke the agent token on the console's Tokens page. Every later call with it gets 401 unauthorized.
• Transformer: revoke the workspace key from the workspace page, or the agent's sign-in from Settings → Agents. Calls with a revoked key get 401.

This is quick only if each agent had its own key. A shared key means cutting off everything that uses it.

What should you set up before anything goes wrong?

• One key or token per agent, scoped to the job, so you can revoke one without stopping the rest.
• A run id on every change, so you can find everything one run did.
• A flag around every change an agent ships, so it can be turned off without a deploy.
• Deadlines and spending limits enforced by the service; see how to stop an AI agent from overspending.
• Decisions that are hard to undo reserved for a person; see how the studio uses agents.