Definition
Plain language
Designing AI agents so their reliability stays good over weeks and months of use, not just on day one.
As stated in the literature
The design discipline of monitoring and intervening on multi-mechanism agent aging across deployment, treating reliability as a lifespan property of the agent-plus-memory system rather than a model snapshot.
Also called: agent lifespan engineering
Why it matters: Without it, agents that pass acceptance tests can quietly degrade in production and no one knows why until users complain.
For example, instead of only benchmarking an agent on day one, the team tracks how its task success rate drifts over six months as its memory store accumulates entries.