AgingBench Benchmark Reveals Lifespan Decline in Deployed AI Agents
Researchers introduce AgingBench, a longitudinal benchmark that measures how AI coding agents deteriorate after deployment. The suite isolates four aging mechanisms—compression, interference, revision, and maintenance—and shows that agents can retain surface performance while factual accuracy drops, requiring stage‑specific repairs rather than stronger initial models.