Restarting Your Agent Is Not Repairing It

Earlier this year I came off my bike and ended up with a fracture.

What stayed with me was watching the body repair itself. Bruises faded. Wounds closed. The fracture stabilised and began knitting back together. Not instantly, not perfectly, but continuously.

A broken bone does not cause the organism to discard itself and restore from last Tuesday’s backup. The damage is localised, stabilised, and repaired while everything else carries on.

I spend my working life around systems that do the opposite.

Recovery is not repair

Modern software is very good at recovery by replacement. Process dies, restart it. Container misbehaves, replace it. This is good engineering because most components are disposable.

But if a component enters a bad state, is destroyed, and reconstructed from an earlier snapshot, the service recovered. The component did not. It was replaced.

For a stateless worker, who cares.

A long-lived agent is different. Over time it can accumulate corrections, unfinished work, evidence, commitments, and relationships with external systems that exist nowhere else.

Suppose an agent takes a lease on an external system at 10:05. Its last checkpoint was…