ADIAS framework automates agent design, beating baselines by 25.2%

Automated agent design methods iteratively revise and evaluate AI agent harnesses, but most are candidate-centric: they organize cross-round experience around candidate agents, making repair progress implicit. This leads to inefficient repair targeting, slow consolidation of partial progress, and propagation of ineffective interventions. To solve this, researchers Lekang Jiang, Bohan Tang, Stephan Goetz, and Yiwen Guo propose issue-centric agent optimization in a new arXiv paper. Their framework, ADIAS (Automated Design of Interactive Agentic Systems), carries repair progress forward as an explicit persistent issue state—tracking issue identities, lifecycle status, supporting evidence, and intervention-outcome histories. An issue-guided optimization mechanism then uses this state to jointly propose repair targets and revision directions for focused, full-code modification, instead of re-deriving context from candidate history every round.

ADIAS was evaluated across five interactive benchmarks, where it outperformed the strongest baseline by 25.2% on average and achieved consistent gains across four different backbone models. Controlled ablations confirmed the value of its design: removing the persistent issue state or replacing issue-centric revision with candidate-centric policies led to performance drops of up to 40.7%. These results suggest that explicitly maintaining problem state throughout the optimization loop is significantly more effective than the candidate-centric paradigm commonly used in prior automated agent design work. For AI engineers, ADIAS represents a more efficient and reliable approach to automatically building and refining agentic systems, potentially reducing wasted repair cycles and enabling faster iteration on complex agent behaviors.