Verified Agentic AI (VAAI): A Multi-Agent Framework for Trustworthy Large Language Models

erified Agentic AI (VAAI): A Multi-Agent Framework for Trustworthy Large Language Models Large Language Models (LLMs) are increasingly being deployed in high-stakes domains such as healthcare, finance, cybersecurity, and critical infrastructure. While recent advances in Agentic AI enable autonomous planning, reasoning, and decision-making, current systems remain vulnerable to hallucinations, reasoning inconsistencies, behavioural instability, and operational risks. Existing approaches—including Retrieval-Augmented Generation (RAG), Knowledge Graphs, self-verification techniques, Explainable AI (XAI), and AI safety methods—primarily address individual aspects of trustworthiness and do not provide a unified verification mechanism prior to decision execution. This paper proposes the Verified Agentic AI (VAAI) Framework, a novel multi-agent verification architecture that introduces an independent verification layer between the Agentic AI reasoning engine and the decision execution layer. The framework comprises four specialised verification agents: Evidence Verification Agent, Behaviour Analysis Agent, Mechanistic Analysis Agent, and Risk & Trust Assessment Agent. Their outputs are integrated by a Trustworthiness Evaluation Engine (TEE), which computes an adaptive Trust Score, generates transparent explanations, and determines whether an AI-generated decision should be approved, subjected to additional verification, or escalated for human review. To demonstrate the feasibility of the proposed framework, a prototype implementation is evaluated using a financial case study based on Sri Lankan inflation forecasting. Responses generated by multiple state-of-the-art Large Language Models (ChatGPT, Claude, Gemini, and DeepSeek) are processed through the complete verification pipeline, including claim extraction, evidence verification, behavioural analysis, mechanistic reasoning assessment, risk evaluation, and adaptive trust score computation. The proposed framework provides a foundation for trustworthy, explainable, and verifiable Agentic AI systems capable of supporting reliable autonomous decision-making across diverse high-risk application domains.