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

EngiAgent: Fully Connected Coordination of LLM Agents for Solving Open-ended Engineering Problems with Feasible Solutions

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent collaboration and safety · 0 citations (OpenAlex)

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 EngiAgent, a five-agent LLM system with a fully connected coordinator that routes feedback among analyzer, modeler, verifier, solver and evaluator, and evaluates feasibility of numerical solutions to 53 open-ended engineering problems.

Problem

LLMs often produce engineering solutions that look plausible or are numerically computed but violate data consistency, physical laws, or constraints, so feasibility of open-ended engineering modeling is rarely ensured.

Approach

Replaces fixed pipeline reflection with a fully connected coordinator that dynamically routes error feedback among specialized agents (analysis, modeling, verification, solving, evaluation), plus forced agent switching to prevent repeated debugging loops; feasibility is judged by human experts rather than LLM judges.

Key results

• EngiAgent (Coord.) reaches 64.15% feasibility on GPT-4o versus 7.55% for the best baseline (MM-Agent).
• On DeepSeek-V3-671B, EngiAgent (Coord.) achieves 75.47% feasibility versus 28.30% for the best baseline (DS-Agent).
• The fully connected coordinator improves feasibility over the fixed pipeline by 16.98 points on GPT-4o and 11.32 points on Gemini-2.5 Flash.

Surprising findings

• Text-based LLM-style scores are inflated: many infeasible solutions still receive high textual ratings, so feasible and infeasible score distributions overlap.
• Zero-shot and ResearchAgent produce almost no feasible solutions on Gemini-2.5 Flash and GPT-4o, while DS-Agent reaches 60.38% numerical solutions on Gemini but 0% feasible.

Limitations

• Benchmark is small (53 problems), limiting statistical power.
• Feasibility labels depend on expert judgment and some annotator disagreement (4 of 53) was observed.
• Higher runtime and token usage than some baselines, e.g., up to 1,743 s on Gemini for the fixed pipeline.

Experimental setup

Datasets: EngiAgent benchmark (53 engineering problems from peer-reviewed papers)

Benchmarks: EngiAgent benchmark

Metrics: feasibility rate numerical solution rate IE/DR/MO/UH scores (0-10) cost and runtime

Baselines: Zero-shot prompting ResearchAgent DS-Agent MM-Agent EngiAgent (Fixed)

Builds on: EngiBench ResearchAgent DS-Agent MM-Agent

Closest papers by meaning

• MM-Agent: LLM as Agents for Real-world Mathematical Modeling Problem (NeurIPS 2025 · similarity 0.87)
• AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.86)
• Lessons Learned: A Multi-Agent Framework for Code LLMs to Learn and Improve (NeurIPS 2025 · similarity 0.85)
• PathWise: Planning through World Model for Automated Heuristic Design via Self-Evolving LLMs (ICML 2026 · similarity 0.85)
• Toward Engineering AGI: Benchmarking the Engineering Design Capabilities of LLMs (NeurIPS 2025 · similarity 0.85)
• SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents (NeurIPS 2025 · similarity 0.85)