TeLLAgent: a dual-agent framework for reliable scientific discovery with tool-enhanced LLMs

Area of Science:

• Artificial intelligence
• Materials science
• Scientific discovery

Background:

• Large language model (LLM) agents show potential for scientific discovery but struggle with reliable tool orchestration and complex planning, leading to hallucinations and inconsistencies.
• Existing agent frameworks lack the robustness required for real-world scientific automation.

Purpose of the Study:

• To introduce TeLLAgent, a novel supervisor-executor dual-agent framework designed to overcome limitations in LLM agent reliability for scientific discovery.
• To enhance the ability of AI agents to orchestrate tools, execute multi-step plans, and minimize logical inconsistencies and hallucinations.

Main Methods:

• TeLLAgent employs a dual-agent architecture: a global planning agent (DeepSeek-R1) for iterative reasoning and dynamic plan formulation, and a local execution agent (DeepSeek-V3.1) for precise tool invocation.
• A self-correction loop, mediated by the Model Context Protocol, enables the system to "rethink" and "recover" from execution failures, enhancing robustness.
• The framework was benchmarked on complex tool-calling tasks and demonstrated on an end-to-end organic solar cell material discovery workflow.

Main Results:

• TeLLAgent significantly outperformed GPT-5 and existing agent frameworks in success rates for multi-step planning and scalability with task complexity.
• The system drastically reduced factual hallucinations in knowledge retrieval, validated by human experts and LLM judges.
• An AI-designed quasi-macromolecular acceptor molecule was identified, synthesized, and validated, achieving a 16.44% power conversion efficiency in organic solar cells.

Conclusions:

• TeLLAgent establishes a new paradigm for reliable, autonomous AI systems capable of accelerating scientific discovery.
• The supervisor-executor dual-agent framework with self-correction mechanisms significantly enhances LLM agent performance and reliability.
• TeLLAgent demonstrates practical application in materials science, with potential for broader impact in fields like drug discovery.