Human-Agent Collaboration Mechanism for Software Testing

Bachelor & Master Thesis

This research proposes a principled collaboration mechanism that lets human testers and LLM-based agents work as coordinated peers across the testing lifecycle. The mechanism defines (1) role allocation: humans contribute domain intent, risk heuristics, and acceptance criteria; agents generate test ideas, execute scenarios, and summarize evidence; (2) a shared artifact space: structured test charters, traceable requirements, and executable test plans maintained with provenance; and (3) interaction protocols: turn-taking prompts, critique loops, and negotiation rules when agent recommendations conflict with human judgment. Agents expose confidence, assumptions, and coverage deltas; humans provide corrective feedback that updates agent policies via lightweight reinforcement signals. The system includes guardrails for data privacy, deterministic replay for audits, and escalation paths for ambiguous failures. Expected outcomes are measurable quality gains, transparent decision trails, and a repeatable blueprint for human–AI testing teams.

Required knowledge: python programming and data processing, experience with CV/NLP and (multimodal) LLMs, web crawling techniques, experience with mining software repositories, experience of automated GUI testing techniques.