Optimize agent systems against a persistent issue list, not against candidate history — 25.2% over the best baseline, 40.7% drop without it

arXiv 2608.06410·medium signalXBlueskyLinkedInCopy link

ADIAS reframes automated agent-system design as issue-centric: instead of re-deriving what's broken from the previous round's candidates each iteration, it carries an explicit persistent issue state holding stable issue identities, lifecycle status, supporting evidence, and the history of which interventions helped. That state drives targeted full-code repairs and delivered a 25.2% average improvement over the strongest baseline across five interactive benchmarks and four backbone models, with ablation showing performance falls up to 40.7% when the issue state is removed. Directly portable to any self-improving harness that currently re-diagnoses its own failures from scratch every cycle.