Martin Fowler's team published a detailed look at how infrastructure can keep agents running across hours or days without losing work to failures. The core idea: record each step as it completes, so a crashed agent can pick up from the last safe point rather than start over. This changes what a human sees. A long-running task stops feeling like a black box and starts feeling like something with a recoverable history.
For a designer, the interaction question shifts from 'did it finish?' to 'where did it stop, and can I see why?' The piece argues that agents doing real work need infrastructure that makes failures visible and reversible. A human reviewing a failed run needs to see which step broke, what state the agent held at that moment, and whether resuming is safe. That is a progress and recovery problem as much as an engineering one.
The piece stays close to the engineering layer, so design implications require reading between the lines. The strongest signal for product teams is this: the longer an agent runs, the more the human relationship shifts from delegation to monitoring, and the more the product must surface what the agent has done so far, not just what it plans to do next.