Glonce

ADIAS beats agent-design baselines by 25.2% on average

researchThe Glonce editorsarXiv🔥70Glonce rating70Impact68Novelty72Relevance95Surprise42Players25Credibility82

Automated agent design improves agent harnesses through repeated cycles of revision, evaluation and feedback summarization. The authors argue that existing methods are largely candidate-centric: they organize what was learned across rounds around the candidate agents themselves, which leaves the actual repair progress implicit. They say this causes three problems: inefficient targeting of what to repair, slow consolidation of partial progress, and the propagation of interventions that did not work into later rounds.

To address this, the authors formulate issue-centric agent optimization, an approach in which repair progress is carried forward as an explicit persistent issue state that guides optimization, rather than being re-derived from candidate history in every round. They instantiate this idea in ADIAS, a framework for automated full-code agent design built on two mechanisms. The first, a persistent issue state, maintains stable issue identities, their lifecycle status, supporting evidence, and the history of which interventions were tried and what they produced. The second, issue-guided optimization, uses that state to jointly propose what to repair next and in which direction to revise, feeding into focused full-code modification rather than broad rewrites.

Across five interactive benchmarks, ADIAS outperforms the strongest baseline by 25.2% on average, with consistent gains across four backbone models. Controlled ablations show the persistent issue state is doing real work: removing it, or swapping issue-centric revision for a candidate-centric policy, drops performance by as much as 40.7%. The source text does not name the five benchmarks, the four backbone models, the authors, their institutions, or a submission date.

Key facts

  • ADIAS reformulates automated agent design as issue-centric optimization: repair progress is kept as an explicit persistent issue state instead of being re-derived from candidate history each round.
  • The persistent issue state tracks stable issue identities, lifecycle status, supporting evidence, and the history of which interventions were tried and their outcomes.
  • Issue-guided optimization uses that state to jointly pick repair targets and revision directions for focused full-code modification.
  • Across five interactive benchmarks, ADIAS beats the strongest baseline by 25.2% on average, with consistent gains across four backbone models.
  • Ablations show removing the persistent issue state, or replacing issue-centric revision with a candidate-centric policy, drops performance by up to 40.7%.

Why it matters

Automated agent design tools iterate a coding agent through cycles of revision and evaluation to fix its weaknesses. The authors' diagnosis is that most such tools track this process by candidate agent rather than by the underlying problem, so partial progress on a specific issue gets scattered across rounds instead of accumulating, and fixes that already failed can get tried again. Making the issue itself, not the candidate, the persistent unit of memory is the core shift ADIAS proposes, and the paper's ablations are offered as evidence that this shift is what drives the reported gains rather than the full-code revision mechanism alone.

Who it affects

The work targets people building or evaluating automated agent design and agent-harness optimization systems: researchers working on self-improving coding agents and the interactive benchmarks used to measure them. It does not describe an end-user product or a general audience.

How to use it

ADIAS is presented as a research framework and method, not a released product or tool. The source text gives no code availability, pricing, licence terms, or named benchmarks and backbone models, so there is nothing concrete here about deploying it beyond the method description itself.

How solid is it

The evidence is the authors' own evaluation across five interactive benchmarks and four backbone models, where ADIAS outperforms the strongest baseline by 25.2% on average. The paper backs the mechanism with controlled ablations: removing the persistent issue state, or replacing issue-centric revision with a candidate-centric policy, cuts performance by up to 40.7%, which is offered as evidence the persistent issue state itself is doing the work rather than some other part of the pipeline. The source text does not name the benchmarks or backbone models used, nor the authors or their institutions, which limits independent verification.

Risks and caveats

The claims come from the authors' own benchmarking, with no third-party replication described in the source. Neither the five benchmarks nor the four backbone models are identified in the text, so the generality of the 25.2% average gain across different task types or model families cannot be checked from what is given. No author names, institutional affiliations, or submission date are stated, and no information is given about code release or real-world deployment.