AgingBench: A Benchmark for Measuring AI Agent Performance Degradation

A new benchmark called AgingBench has been released to measure the phenomenon of AI agents' performance degrading over time after deployment. This research focuses not on single-task performance, but on how agents hold up in long-term deployment environments.

Key findings of the research are as follows:

• The paradox of model replacement: In the Claude Code CLI environment, when the backbone model was replaced from Sonnet 4.6 to Opus 4.7, the PyTest pass rate actually dropped by an average of 15%.
• Evolution of memory state: Performance degradation is not a problem with the model's underlying capability, but rather a longitudinal effect caused by changes in memory state during long-term operation, such as memory compression, interference, modification, and maintenance shocks.
• The importance of memory policy: Memory policy was shown to have a greater impact on agent half-life than model replacement (a 4.5x difference).

In conclusion, for long-running agent systems, simply switching to the latest model may not be a safe upgrade strategy, and memory management policy plays a central role.