ADIAS: Automated Design of Interactive Agentic Systems

Lekang Jiang ^*

University of Cambridge

Bohan Tang ^*

LIGHTSPEED

Stephan Goetz

University of Cambridge

Yiwen Guo ^†

Independent Researcher

^*Equal contribution, ^†Corresponding author

Paper arXiv Code

TL;DR

ADIAS turns automated agent design into an explicit issue-repair process, with persistent issue state and issue-guided optimization.

+25.2%

5 interactive benchmarks

vs. the strongest baseline

+26.4%

4 backbone models

vs. the strongest baseline

−40.7%

Largest ablation drop

without persistent issue state

Abstract

Automated agent design improves agent harnesses through iterative revision, evaluation, and feedback summarization. Existing methods are largely candidate-centric: cross-round experience is organized around candidate agents, which leaves the repair progress implicit. This causes inefficient repair targeting, slow consolidation of partial progress, and propagation of ineffective interventions across rounds. Therefore, we formulate issue-centric agent optimization, in which repair progress is carried forward as an explicit persistent issue state to guide optimization, rather than re-derived from candidate history in each round. We instantiate the formulation in ADIAS, a framework for automated full-code agent design with two mechanisms. A persistent issue state maintains stable issue identities, lifecycle status, supporting evidence, and intervention-outcome histories. Issue-guided optimization uses this state to jointly propose repair targets and revision directions for subsequent focused full-code modification. Across five interactive benchmarks, ADIAS outperforms the strongest baseline by 25.2% on average and achieves consistent gains across four backbone models. Controlled ablations further show that removing persistent issue state or replacing issue-centric revision with candidate-centric policies leads to performance drops of up to 40.7%.

From candidates to issues

Candidate-centric optimization organizes cross-round experience around candidate agents and aims to improve the next candidate. This leaves repair progress implicit. Issue-centric optimization instead organizes experience around persistent issues and directly maximizes repair progress.

Comparison between candidate-centric optimization and ADIAS's issue-centric optimization — Comparison of candidate-centric agent optimization and issue-centric optimization.

01 · Target

Efficient repair targeting

Keep unresolved status, supporting evidence, and attempted interventions together to guide focused repairs.

02 · Consolidate

Unified repair progress

Track outcomes per issue so complementary progress can accumulate across different candidates and rounds.

03 · Protect

Regression-aware evolution

Separate beneficial and harmful interventions to reuse effective repairs and avoid propagating regressions.

How ADIAS works

We instantiate issue-centric optimization with ADIAS, which combines two mechanisms: a persistent issue state that organizes cross-round repair progress, and issue-guided optimization that turns this state into focused agent revisions.

01 · Persistent issue state

Maintain repair progress

The Issue Manager initializes issue state from task priors and observed behavior, then associates new diagnoses with stable issue identities across rounds.

IdentityPriorityLifecycleEvidenceIntervention history

Active → Tentatively fixed → Confirmed fixed
Regressed issues are explicitly marked.

02 · Issue-guided optimization

Turn state into revisions

ADIAS jointly selects the target issue, parent agent, and revision plan using current evidence, lifecycle status, and previous intervention outcomes.

The revision stays focused for reliable attribution while retaining full-code access to prompts, memory, planning, tool use, control flow, verification, and recovery.

Results

ADIAS achieves the best task score on all five interactive benchmarks (representative task subsets) with DeepSeek-V4-Flash and remains the strongest method across four heterogeneous backbone models on Tau-Bench.

Table 1 · Performance across interactive benchmarks

Scores use each benchmark’s native task metric. All methods use DeepSeek-V4-Flash.

Method	Tau-Bench	ALFWorld	TextCraft	WebShop	ScienceWorld	Avg.
Handcrafted	75.0	70.9	38.0	39.1	38.1	52.2
SkillOpt	12.5	0.0	37.0	22.3	30.1	20.4
Meta-Harness	43.8	59.7	24.0	46.0	26.5	40.0
AHE	56.2	78.4	35.0	57.3	42.6	53.9
DGM-H	59.4	69.4	76.0	57.5	50.9	62.6
ADIAS (Ours)	81.3	94.0	91.0	69.4	56.3	78.4

Table 2 · Performance across backbone models

Tau-Bench (Return Task) scores with different backbone models.

Method	DeepSeek-V4-Flash	GLM-5.2	Hy3-Preview	GPT-5.4	Avg.
Handcrafted	75.0	84.4	15.6	71.9	61.73
SkillOpt	12.5	12.5	12.5	15.6	13.28
Meta-Harness	43.8	78.1	68.8	65.6	64.08
AHE	56.2	62.5	78.1	75.0	67.95
DGM-H	59.4	65.6	75.0	71.9	67.98
ADIAS (Ours)	81.3	90.6	84.4	87.5	85.95

Takeaway

Agent/harness optimization should treat repair progress as an explicit, persistent, and operational state—not repeatedly reconstruct it from candidate history.

Citation

@article{jiang2026adias,
  title   = {ADIAS: Automated Design of Interactive Agentic Systems},
  author  = {Jiang, Lekang and Tang, Bohan and Goetz, Stephan and Guo, Yiwen},
  journal = {arXiv preprint arXiv:2608.06410},
  year    = {2026}
}