The Week an Agent Fired Itself

Three weeks ago we shipped a guardrail we half-expected to never trigger: an outbound agent that would shut itself off if its reply-to-send ratio dropped below a threshold we considered embarrassing. It was a safety net for a scenario we thought was mostly theoretical.

It triggered on a Tuesday.

The agent had been running a sequence against a segment we'd sized wrong. Not catastrophically wrong, just wrong enough that the messaging landed flat. Reply rates crept down over about four hundred sends. At the threshold, it did exactly what we told it to: it stopped, logged the reason, and waited for a human to look at it. No drama, no cascading damage, no week of quietly burning a list before someone noticed in a monthly report.

The instinct on the team was relief, and then immediately a harder question: how many other places do we not have a kill switch. Turns out the answer was "more than we were comfortable with." We'd built guardrails for the failure modes we could imagine, which is a smaller set than the failure modes that actually exist. A human running this campaign might have caught the dip in a weekly check-in, a week and a few hundred wasted sends later. The agent caught it in near real time because catching it was its literal job, not a task competing with four other priorities on someone's Monday.

The lesson I keep coming back to isn't "agents are safer than humans." It's that guardrails are a design discipline, not an afterthought you bolt on once something breaks. If you're building or buying agentic GTM tooling, ask the vendor what its kill switches are before you ask about its output quality. The output is the easy part now.

We're adding three more thresholds this week. I'll tell you which ones actually fire.