AI Agents Need to Be Replaceable Without Losing Your Work
The New Stack, Wednesday, October 28th, 2026
When AI agents crash mid-task, durable records and tool boundaries let replacement agents resume work safely without duplicate actions.
The New Stack looks at what happens when an AI agent fails partway through a task and a replacement has to pick up the work.
In its example, an agent approved to issue a partial refund calls a payment API and crashes, leaving an operation record marked dispatched but no outcome, so nobody knows whether the refund went through.
Any distributed worker that crashes mid-change faces this ambiguity, but while a conventional worker resumes a fixed workflow from its last checkpoint, an agent reasons on its own and might choose different parameters or issue a new refund instead.
The article argues durable operation records and tool boundaries are what let replacement agents resume safely.