Research map / Multi-agent systems and games / LLM agents

Discovering Hierarchical Software Engineering Agents via Bandit Optimization

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes BOAD, which casts automatic discovery of hierarchical SWE sub-agent teams as a multi-armed bandit problem with LLM-judged hindsight helpfulness credit, achieving 20.0% on SWE-bench Live and 53.2% on Verified with a 36B model.

Problem

Single-agent SWE systems generalize poorly to long-horizon, out-of-distribution issues, and automatically designing multi-agent hierarchies is combinatorial and suffers from ambiguous per-sub-agent credit assignment.

Approach

Treat each candidate sub-agent as a bandit arm, select top-K via UCB, expand the archive with a Chinese Restaurant Process, and score each sub-agent by an LLM judge's binary helpfulness label over trajectories (hindsight credit) rather than team success rate.

Key results

• On SWE-bench Live, BOAD with Seed-OSS-36B-Instruct resolves 20.0% (60/300) vs 12.3% for the SWE-agent baseline and 17.0% for evolutionary search.
• On SWE-bench Verified, BOAD reaches 53.2%, versus 49.8% for the SWE-agent baseline and 47.4% with manually designed sub-agents.
• Ablation shows top-2 sub-agents peak at 20.0% while top-5 drops to 13.7%, and helpfulness-based selection beats success-rate selection (20.0% vs 15.3%).

Surprising findings

• Manually designed sub-agents from prior work lower performance on SWE-bench Verified (47.4% vs 49.8% baseline).
• Larger teams of sub-agents (3 to 5) reduce performance, with exactly two sub-agents performing best.

Limitations

• Discovered sub-agents transfer only partially to other models (Claude 3.7 Sonnet gains smaller).
• Error propagation: orchestrator accepts unvalidated sub-agent outputs.
• Evaluated only on a small 12-issue design set and modest iteration budget (B=20).

Experimental setup

Datasets: SWE-bench Verified (12-issue design set) SWE-bench Live

Benchmarks: SWE-bench Verified SWE-bench Live

Metrics: resolved rate (%) total tokens max input tokens

Baselines: SWE-agent (default) SWE-agent + Manual Sub-agent SWE-agent + Evolutionary Search GPT-4o Claude 3.7 Sonnet DeepSeek-V3 GLM-4.5-Air Qwen3-Coder-30B-A3B-Instruct

Builds on: SWE-agent Automated Design of Agentic Systems UCB Chinese Restaurant Process

Closest papers by meaning

• SWE-Search: Enhancing Software Agents with Monte Carlo Tree Search and Iterative Refinement (ICLR 2025 · similarity 0.91)
• SWE-rebench: An Automated Pipeline for Task Collection and Decontaminated Evaluation of Software Engineering Agents (NeurIPS 2025 · similarity 0.89)
• AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.89)
• Kimi-Dev: Agentless Training as Skill Prior for SWE-agents (ICLR 2026 · similarity 0.88)
• GSO: Challenging Software Optimization Tasks for Evaluating SWE-Agents (NeurIPS 2025 · similarity 0.87)
• Diversity Empowers Intelligence: Integrating Expertise of Software Engineering Agents (ICLR 2025 · similarity 0.87)