Study planning assistant: Supporting material discovery with agentic workflows

Agentic Workflow refers to an innovative agents-based AI approach that assists users on achieving complex goals by decomposing them into task sequences. In this work, we present the Study Planner Assistant (SPA), a LLM-based agentic workflow that supports subject matter experts to translate their goals into actionable activities given the current available tools in a computational system.

The SPA provides an UI to configure, inspect and control the workflow (see figure) and a backend composed of four main modules:

1. Plan creation agent: takes as input the user goal and a collection of available tasks to generate the plan (tasks sequence).
2. Tasks allocation agent: responsible for allocating tasks, given the user preferences, to the proper tools that can be executed.
3. Workflow generation agent: generates the code to call the tools in a sequence respecting the order defined in the plan and makes the output of a tool compatible with the following tools.
4. Workflow execution: coordinates the execution of the different tools.

The first 3 modules are powered by LLMs. Each LLM is prompted with the module's input and the expected output format, guaranteeing the integration between modules. The last module uses LangGraph to execute the workflow and coordinate the information between the tasks.

In this work, we will present how the SPA was applied to a Material Discovery use case (PFAS replacement) and demonstrated great potential.