Δτ-Bench tests AI agents on changing tools, with STEWARD ecosystem manager

Ningyu Zhang introduced Δτ-Bench, a benchmark that tests whether AI agents keep pace with evolving tool ecosystems, where APIs change, tools disappear, and better alternatives appear.

The authors report that models struggle to reach the current frontier and that skill-based methods do not consistently hold up through ecosystem changes.

The authors also introduce STEWARD, an ecosystem manager that maintains tools and skills together. They report better task success and lower maintenance costs than the compared approaches. The source for these results is the launch announcement; the underlying paper's figures were not included in the report.