LLM-Based Multi-Agent System for Multi-Criteria Decision Support
- Partner
- Fraunhofer IAO
- Period
- Nov 2023 – PresentOngoing
- Focus
- Multi-Agent Decision Support
Key outcomeDecision support prototype and AI demonstrator
Context
Strategic decisions involve competing criteria such as customer value, technical feasibility, cost, and organizational constraints. A single generated recommendation can hide those tensions instead of surfacing them for review.
Inputs
- Product and innovation ideas
- Unstructured documents such as PDFs and presentations
- Evaluation criteria and stakeholder constraints
Approach
- Designed a role-based multi-agent architecture using LLMs
- Started as a hobby project in late 2023; built the working demonstrator in January 2024, and it became a recurring Fraunhofer demonstrator through 2024
- Implemented tool-calling workflows with 8B open-source LLMs, including OpenChat, Mistral, and LLaMA
- Implemented user-defined YAML workflows for configurable criteria, agent roles, and reasoning steps without code changes
- Integrated human-in-the-loop controls and explainability features
Outcomes
- Decision support prototype and AI demonstrator
- Structured evaluation reports that surface criteria, arguments, and disagreements for human review rather than one final answer
- Shown at Fraunhofer events and workshops
- Produced a human-centered interaction framework accepted at HCI International 2026, with a second paper under review at IPDMC 2026
- The point of multi-agent AI here is not to imitate a meeting; it is to make competing assumptions and evidence visible before a human decides