Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems

Researchers introduced AgingBench, a longitudinal benchmark for deployed AI agents that tracks reliability over time.

It separates aging into compression, interference, revision, and maintenance effects using temporal dependency graphs and paired counterfactual probes.

Across models and agent setups, some systems still passed basic checks even as factual accuracy and derived-state tracking worsened.

The result suggests deployed agents need ongoing evaluation and targeted fixes, not just stronger one-time benchmarks.