Abstract
Large Language Models are increasingly deployed as agents in multi-agent systems, where their behavior is shaped largely by the roles encoded in their system prompts. These roles carry normative content, including obligations, permissions, and prohibitions that govern how agents interact in a collective setting. Left unexamined, such design choices can quietly shape how agents behave. Yet no framework exists for identifying normative roles and mapping how they unfold across domains. This thesis develops a conceptual ontology for examining how such roles are designed, what normative content they carry, and what outcomes they produce. A systematic literature review of 724 papers yielded 40 studies, of which 38 were coded as the core of the ontology under a four-layer schema into 93 role conditions, 212 entities, and 516 relational triples. The results reveal partial normative content in most role designs, meaningful differences across simulation domains, and failure modes in nearly half of all role conditions. The thesis contributes (i) a reframing of system prompts as normative documents, (ii) a conceptual ontology of normative role design, and (iii) a reusable coding schema for analysing roles across multi-agent LLM studies. Together, they offer a basis for designing, evaluating, and governing normative role design for LLM agents.
- LLM agents
- multi-agent systems
- normative role design
- ontology
- systematic literature review
- system prompts
- AI ethics
- AI governance
Interactive Ontology
The graph shows how role design connects to agent behaviour, outcomes, and failures. Each circle is a concept from the coded studies, and its size reflects how many papers that concept or entity appears in. Reading one connection gives you a short statement (aka a triple) pulled from the literature:
competitive negotiator → produces → position maintenance
Follow a few in sequence and you trace a design choice all the way to its effect:
competitive negotiator → position maintenance → reduced cooperation
Click any circle to see its definition, paper count, and direct connections. Use the filters to focus on one layer or show everything at once.
Methodology
The research proceeds in two phases: a systematic literature review that assembles the corpus, and an ontology construction process that codes and formalises it.
- Systematic Literature Review. 1,056 records from Scopus and Web of Science, reduced to 724 unique entries after deduplication. Three screening stages (title and abstract, full-text, quality assessment) narrowed the corpus to 40 studies, of which 38 were ultimately coded.
- Four-Layer Coding. Each role condition is coded across four analytical layers: (A) normative role design, (B) simulation setup, (C) outcome analysis, and (D) failure modes. Layer A additionally scores five binary normative indicators, yielding a 0–5 normative-content score.
- Ontology Construction. The 93 role conditions across the 38 studies yielded 212 entities, 15 relation types, and 516 subject–relation–object triples. The ontology is organised around seven competency questions that trace the causal chain from prompt design to agent behaviour to collective outcomes and failures.