ADIAS Automates Agent Design with Issue-Centric Optimization.

Current methods for automating the design of AI agents often struggle with efficiency because they focus on individual candidate agents, making it hard to track and consolidate repair progress. This new framework, ADIAS, proposes an issue-centric approach. Instead of re-deriving problems from scratch for each agent candidate, ADIAS maintains a persistent state for each identified issue, including its lifecycle, evidence, and intervention history. This allows for more targeted and efficient modifications to the agent's code. The ADIAS framework uses this issue-guided optimization to jointly suggest repair targets and directions for code changes. Evaluated across five interactive benchmarks, ADIAS significantly outperformed existing baselines, showing an average improvement of 25.2%. Ablation studies confirmed that the persistent issue state and issue-centric revision are critical to its superior performance.