- The paper introduces EngiAgent, a multi-agent framework using dynamic coordination that achieves up to 7x improvement in feasibility over traditional approaches.
- The system decomposes the engineering workflow into specialized agents that collaboratively ensure strict adherence to operational, physical, and safety constraints.
- Experimental evaluations across diverse engineering domains demonstrate significant reductions in error rates and improved practical implementability of solutions.
EngiAgent: A Fully Connected Multi-Agent LLM Framework for Feasibility-Centric Engineering Problem Solving
Motivation and Context
Solving real-world engineering problems demands more than mathematical formulation or code synthesis; it requires producing solutions that are feasible with respect to domain-specific operational, physical, and safety constraints. Despite LLM advances in reasoning and code generation, empirical evidence indicates that feasibility—i.e., the enforceability of constraints and implementability in practice—is consistently neglected. For example, state-of-the-art LLM-based agents achieve as little as 3.77–7.55% feasibility across engineering tasks, even when their numerical output rates are much higher. This gap arises because most existing approaches focus on correctness or expressiveness in modeling, rather than ensuring that generated solutions satisfy all real-world requirements.
The paper "EngiAgent: Fully Connected Coordination of LLM Agents for Solving Open-ended Engineering Problems with Feasible Solutions" (2605.02289) introduces EngiAgent, a multi-agent system featuring a fully connected coordination mechanism explicitly designed for end-to-end feasibility assurance. EngiAgent embodies the expert engineering workflow through modular, specialized LLM agents while dynamically orchestrating inter-agent feedback to correct failures and adaptively route error recovery operations. This framework establishes a foundational architecture for constraint-driven engineering automation with LLMs.
Feasibility as the Central Criterion
The core insight driving EngiAgent is that feasibility is non-negotiable in engineering tasks. Unlike mathematical problem-solving, where logical correctness implies validity, engineering solutions face multi-stage breakdowns: data extraction errors, misapplied or missing constraints, code execution failures, and physical law violations. The framework defines feasibility as strict satisfaction of all mandatory engineering constraints—physical, operational, data, and safety requirements—that guarantee the solution's ability to function in real systems. Even minor violations result in infeasibility, regardless of apparent numerical or logical plausibility.
The authors categorize typical LLM-driven errors as:
- Fancy but vague modeling (i.e., sophisticated language with lack of concrete executable equations or variables)
- Data alteration (modifying essential parameters for solvability, thereby deviating from the intended problem)
- Physical law violation (e.g., exceeding device capacities, disregarding conservation)
- Over-constraining (introducing superfluous requirements that preclude any feasible solution)
These failure modes are not mitigated by current pipeline-based or single-agent systems, motivating an architecture capable of adaptive error diagnosis and correction at every stage.
System Architecture
EngiAgent decomposes the engineering workflow into five functionally specialized agents:
- Analyzer: Transforms unstructured problem statements into structured, context-aware model blueprints, extracting variables, parameters, and constraints, and identifying implicit domain rules.
- Modeler: Translates problem blueprints into executable, domain-correct code, with built-in revisions for semantic or specification mismatches.
- Verifier: Conducts hierarchical, pragmatic verification for semantic consistency, constraint completeness, and data fidelity, triggering targeted error correction rather than generic rejection.
- Solver: Executes models with automated solver selection and resource management, feeding back detailed status and error types.
- Evaluator: Aggregates holistic feedback, scoring solution feasibility, domain alignment, and robustness, and directing further iteration where necessary.
Central to EngiAgent is the Fully Connected Coordinator, which supports non-sequential, memory-augmented, and error-aware routing across all agents. This coordinator leverages forced agent switching, error-state history, and process convergence rules. When a feasibility or execution error is detected at any stage, the coordinator can direct correction not only to the next agent in a linear pipeline but to any responsible role. It thereby addresses error propagation, debugging loops, and inter-agent misalignment that routinely defeat fixed-pipeline approaches.
Experimental Evaluation
A new benchmark was compiled spanning four domains—market/multi-agent decision-making, scheduling/resource allocation, planning/design, and control/autonomous system modeling. Each of the 53 high-quality open-ended engineering tasks is annotated for essential feasibility constraints by expert consensus. Evaluation focuses primarily on feasibility, with secondary axes including information extraction, domain-specific reasoning, multi-objective trade-off handling, and robustness to uncertainty.
EngiAgent was compared to several SOTA LLM-based problem-solving agents, including zero-shot prompting, ResearchAgent, DS-Agent, and MM-Agent, across three flagship LLM backends (GPT-4o, Gemini-2.5 Flash, DeepSeek-V3-671B). The results are unequivocal: EngiAgent (with full coordination) achieves feasibility rates of 64.15% (GPT-4o), 50.94% (Gemini-2.5 Flash), and 75.47% (DeepSeek-V3-671B), corresponding to a mean 7x improvement over the strongest baseline (7.55%, 3.77%, and 28.30% respectively). Notably, the gap between numerical solution rate and feasible solution rate is nearly eliminated, confirming that EngiAgent's design directly closes the implementability chasm.
An ablation study underscores the critical contribution of the fully connected coordinator: replacing it with a fixed-pipeline schedule reduces feasibility by >10% consistently, while removing the verifier results in a substantial drop in feasibility and downstream correctness. Forced agent switching is shown to limit infinite debugging cycles without degrading success. These findings demonstrate that robust multi-stage error handling and dynamic agent reassignment, rather than more elaborate prompts or isolated agent improvements, are central to feasibility-centric engineering automation.
Cost Efficiency and Extensibility
Despite the added coordination, EngiAgent maintains competitive computational efficiency and resource usage. While runtime and token usage are higher than trivial baselines, EngiAgent's cost per solution remains lower than highly iterative approaches and is proportional to the achieved improvements in feasibility. Furthermore, the architecture is tool-agnostic, facilitating extension to alternative modeling stacks, simulation paradigms, and mixed statistical/physics-based solution approaches. The agent and coordinator prompt schemes are modular, supporting rapid adaptation to domains such as ML-based regression with MILP-driven feature search or direct numerical ML without optimization solvers.
Implications and Future Directions
This work establishes an actionable architecture for LLM-enabled engineering automation, with robust feasibility assurance as a foundational design principle. The framework's modularity and coordination mechanism are general and could be adapted to new engineering domains requiring constraint-rich, multi-objective, and uncertainty-resilient workflows.
Potential theoretical implications lie in bridging the gap between LLM-based reasoning and constraint-based symbolic AI; a principled multi-agent and coordinator approach can serve as the control substrate for broader AI-based engineering co-pilots. Practically, EngiAgent may enable reliable and auditable automation for design, planning, and operational tasks, serving as an assistive tool for expert engineers in high-stakes environments.
Challenges remain around interpretability, safe deployment, and the handling of implicitly underspecified constraints; the system promotes auditing, traceability, and human oversight, but does not obviate the need for domain expert intervention in ambiguous or safety-critical scenarios. Future work is warranted on scaling to larger, more realistic problem sets, dynamic inter-domain workflows, and deeper integration across heterogeneous modeling and simulation frameworks.
Conclusion
EngiAgent demonstrates that fully connected, modular LLM-agent systems with dynamic error-aware coordination are essential for achieving feasible solutions to open-ended engineering problems. By prioritizing feasibility validation at every stage, the framework achieves a dramatic improvement over existing LLM-based approaches without sacrificing efficiency or flexibility. This establishes a technically rigorous foundation for next-generation feasibility-oriented intelligent systems for engineering and allied scientific domains.