Research map / Multi-agent systems and games / LLM agents
AgentSquare: Automatic LLM Agent Search in Modular Design Space
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.
Introduces the Modularized LLM Agent Search (MoLAS) problem and AgentSquare, which searches a four-module (planning, reasoning, tool use, memory) agent design space via module evolution, recombination, and an in-context performance predictor.
Problem
Hand-crafted LLM agents are task-specific and hard to adapt to new tasks; prior agent search works either optimize prompts or whole code, without a standardized modular space for recombining existing agent components efficiently.
Approach
Abstracts 16 published LLM agents into a modular design space with uniform IO interfaces for four modules, then searches it with LLM-driven module recombination, LLM-programmed module evolution (FunSearch-style), and an in-context LLM performance predictor that skips unpromising candidates.
Key results
- AgentSquare with GPT-4o reaches 0.695 on ALFWorld vs 0.567 for the best hand-crafted agent (Self-refine), and 0.607 on WebShop vs 0.519 (HuggingGPT).
- Reported average gain of 17.2% over best known human-designed agents across six benchmarks.
- Ablation on GPT-4o: removing module recombination drops TravelPlanner from 0.583 to 0.280; removing module evolution drops ALFWorld from 0.695 to 0.649.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add module evolution (ablation) | AgentSquare · ALFWorld | success rate: 0.649 -> 0.695 | Helped |
| add module recombination (ablation) | AgentSquare · ALFWorld | success rate: 0.616 -> 0.695 | Helped |
| add module evolution (ablation) | AgentSquare · ALFWorld | success rate: 0.623 -> 0.651 | Helped |
| add module recombination (ablation) | AgentSquare · ALFWorld | success rate: 0.546 -> 0.651 | Helped |
Surprising findings
- Simple random module recombination outperforms prompt-level search (OPRO) in the reported trajectories.
- On GPT-4o, removing module recombination leaves PDDL performance unchanged at 0.669, showing the components' effects are task-dependent.
Limitations
- Design space is limited to 16 literature-derived agents and is not claimed to be comprehensive.
- Module evolution candidates must be evaluated in the real environment because the predictor cannot score unseen modules, which raises cost.
- Search results depend on the GPT-3.5/GPT-4o backbone and on task-specific search budgets.
Experimental setup
Datasets: ALFWorld ScienceWorld PDDL WebShop TravelPlanner M3ToolEval
Benchmarks: WebShop ALFWorld SciWorld M3ToolEval TravelPlanner PDDL
Metrics: task score success rate progress rate micro pass rate
Baselines: CoT CoT-SC Self-refine ToT Step Back Thought Propagation HuggingGPT Voyager Generative Agents DEPS OPENAGI Dilu
Builds on: ADAS FunSearch OPRO Promptbreeder CoT
Closest papers by meaning
- AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.95)
- AgentBench: Evaluating LLMs as Agents (ICLR 2024 · similarity 0.91)
- AgentNoiseBench: Benchmarking Robustness of Tool-Using LLM Agents Under Noisy Condition (ICML 2026 · similarity 0.90)
- AgentBoard: An Analytical Evaluation Board of Multi-turn LLM Agents (NeurIPS 2024 · similarity 0.89)
- AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.89)
- Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies (ICLR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.