NEXT BEST AGENT SELECTION IN AN ADAPTIVE WORKFLOW

Smart overview of the Invention

An adaptive workflow system is designed to select the next best agent in a sequence of tasks aimed at achieving a specific goal. This system utilizes a large language model (LLM) to analyze user utterances, extract goals, and select appropriate agents based on metadata. The process involves scoring candidate agents and applying business policy constraints to determine the most suitable agent to invoke next. This method enhances flexibility and adaptability in workflows, allowing for dynamic sequencing of tasks.

Background

Traditional workflows rely on predefined sequences, which can be rigid and insufficiently adaptable to changing circumstances. Adaptive workflows, on the other hand, focus on achieving goals rather than following a strict sequence of steps. These workflows often utilize conversational interfaces, enabling users to interact with the system through natural language. Agents, which are atomic functions of APIs, are invoked to perform specific tasks within the workflow. The challenge lies in selecting the most appropriate agent to continue the sequence towards the goal.

Technical Approach

The system extracts goals from user utterances using natural language processing techniques. It then leverages an LLM to select candidate agents by analyzing metadata that describes these agents. The LLM scores each candidate agent, and business policy constraints are applied to refine the selection. The chosen agent is invoked to perform actions that further the goal. This approach does not rely on historical data, making it suitable for adaptive workflows where goals and available agents may frequently change.

Implementation

The system includes a computer program product with instructions stored on a computer-readable medium, executed by a processor. It supports a range of goals, such as product support or account access, and identifies target agents for each goal. Agents are classified based on their functionality and application context, using metadata from an agent catalog. The LLM is prompted with this information to generate a list of top-k candidate agents, which are then evaluated to select the next best agent.

Advantages

This method addresses the limitations of existing agent recommendation systems, which often depend on static historical data and deterministic rules. By using an LLM and real-time metadata analysis, the system can adapt to changing workflows and new or modified agents. It ensures that selected agents conform to API requirements and align with business policies, enhancing the quality and adaptability of the workflow management process.