Year: 2026
Type: dissertation
Abstract: Contemporary LLM-based orchestration platforms such as LangGraph and CrewAI enable rapid composition of agent pipelines, but share three architectural weaknesses. They lack a persistent record of peer performance across runs, expose no formal model of agent capabilities and task constraints, and treat compliance and ethics as ad hoc concerns rather than first-class design primitives. Consequently, allocation quality does not improve with experience and pipelines cannot demonstrate structural or regulatory correctness, a limitation made urgent by emerging regulatory regimes such as the EU AI Act and the NIST AI Risk Management Framework, which codify binding and de facto expectations for transparency and human oversight in high-risk AI deployments. This thesis presents a stigmergy-driven Multi-Agent framework for intelligent task orchestration that addresses all three limitations within a single integrated design. Coordination is mediated by digital pheromone markers deposited on a shared semantic knowledge graph, where a Pheromone Engine reinforces successful assignment patterns while allowing stale ones to decay, so that experience accumulates in the environment rather than in any individual agent and emergent optimisation arises without direct inter-agent communication. Task allocation builds on an Improved Contract Net Protocol whose bid scoring is augmented with this accumulated pheromone history. Structural integrity is enforced by a semantic validation layer, while an Ethics-Check Agent provides a domain-agnostic safety gate that screens all pipeline outputs for hallucinated data, unfair bias, sensitive-data leakage, and unsafe approvals before execution proceeds. An extensive controlled evaluation in a commercial warehouse brokerage scenario supports all five hypotheses. The pheromone engine substantially improves task allocation accuracy as experience accumulates, the dual-layer governance achieves near-complete recall of ground-truth violations at a low false-positive rate, and the full framework outperforms an uncoordinated baseline by a substantial margin at modest latency and token overhead. A user perception survey with domain experts and technical evaluators indicates that the system is regarded as suitable for internal first-draft generation, with mandatory human review required before external use. This work demonstrates how swarm intelligence, the Semantic Web, and modern AI orchestration can be fused to create a Multi-Agent platform that is adaptive, auditable, and governance-aware by design. The full implementation is publicly available at https://github.com/evaveli/agentcy.
Author: Veli, Eva
Institution: Universitat Rovira i Virgili
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7163794101