An agent did its job. It read the ticket, decided the customer was blocked, moved the ticket to a new queue, posted a summary in the support channel, updated the account record, opened an engineering issue, and emailed the customer that help was on the way. Every step was right for the ticket it thought it was reading. It was reading the wrong ticket.
A year of recovery posts on this blog, and every one of them pictured a rollback. So did I.
Rollback Was Built for Deploys
Every recovery plan rests on rollback. Deploy goes bad, restore the last good version, learn, widen the guardrail. December’s post on this blog put it flat: decision recovery speed beats decision quality. That still holds.
Rollback works because a deploy changes one thing. The system has a state. The state has a history. Restore points at the previous one and the system goes back. One person can do it before the standup ends.
That is the picture every leader carries into an agent purchase. It is the wrong picture.
An Agent Changes Five Things
That agent changed no state. It took five actions in five systems. Ticketing. Slack. CRM. Jira. Email. Each system recorded its own action and only its own action. The support platform knows a ticket moved. Slack knows a message was posted. The CRM knows a field changed. None of them knows the other four happened, and none of them knows the five belong together.
There is no last good state to restore. There is a ticket in the wrong queue, a message people already read, a record that now says something false, an engineering issue somebody already picked up, and a customer who was told help was coming for a problem they do not have.
Some of those reverse. Some get corrected. The email never comes back.
Recovery here is a list before it is anything else. Everything the agent touched, in order, across every system, before anyone can decide what to reverse and what to explain. The list is the hard part. Reversing is the easy part once you have it.
Almost nothing sold as agentic today can produce that list.
What the Trace Gives You
Every agent platform will show you a trace. Tool called, output returned, next step taken. A trace shows the sequence and says nothing about whether the sequence was wise. The recovery version of that problem is narrower and more urgent. The trace lives inside the agent’s platform. The damage lives in five other platforms. The trace says the agent called the CRM tool. It does not say which record, what the field said before, or who has read it since.
The list gets built by hand. Somebody opens each system and reconstructs the afternoon. It is the same ten-person call that used to follow a bad deploy, except the deploy came with a diff and the agent comes with a conversation log.
The Question to Ask
Before an agent goes live, make sure you can answer this one question.
When it gets something wrong, show me everything it touched.
Not the trace. The list. Every record, in every system, with what it looked like before. Most platforms will answer with the trace. That answer tells you who builds the list when it matters. Your team builds it by hand, across systems that were never asked to know about each other.
Rollback fixes one system. The agent changed five. Recovery speed still beats decision quality. The clock now starts at the list.