FLASH: A Reliable Workflow Automation Agent
LLM-based agents usually have challenges in reliably executing multi-step tasks even with detailed instructions. This is because of the error propagation nature of step-by-step agent models. We have developed FLASH agent to address these challenges through innovative reasoning and reflection processes. In our real-world case studies, FLASH demonstrated superior accuracy in customer support, incident management and business workflow automation.
In the enterprise world, there are massive repetitive but time-consuming workflows that requires manual efforts. For example, a customer service agent may need to gather various data points to determine if a refund case is valid. A finance work may need to generate regular reports by pulling data from various digital platforms, analyzing trends, and compiling information. Using AI agents to automate these workflows, can significantly boost their daily work efficiency and productivity.
While existing LLM-based agent approaches can manage such tasks with detailed instructions that cover both task-specific and domain-specific knowledge, their accuracy remains a concern. Due to the error propagation nature of step-by-step agent models, even if the agent has a high likelihood of accurately following instructions for a single step the overall accuracy can become quite low after multiple steps. For example, for a 5-step task, even with an 85% chance of accuracy following each step’s instruction, the overall accuracy after 5 steps drops to 44%.
We have designed and developed FLASH agent, which enhances the accuracy of instruction following for multi-step complex tasks. The core components in FLASH, status supervision and hindsight integration, ensure the high reliability and accuracy of any workflow automation. Status supervision accesses the current status of workflow execution and can trigger status-dependent instructions, breaking down complex instructions into simpler ones based on the current status. Hindsight integration leverages LLMs to automatically generate hindsight from past failures, enabling the system to improve its performance as more recurring cases are handled. By integrating the hindsight into the agent’s reflection step, it can help to further improve the reliability of workflow execution by preventing issues that have happened before.
You may find technical details of FLASH agent in our paper (link) (opens in new tab)to read how FLASH advances state-of-the-arts with real-world examples and results. We are actively working with various product groups at Microsoft to build more capable and reliable agents with FLASH technology. FLASH agent was featured at Ignite’24 keynote as the product support agent that is powered by MAIA chip (video clip (opens in new tab)).
Demo Video: Check out the video below to see how FLASH agent can help automate a recurring customer support workflow and improve the productivity of the support agents. FLASH agent reliably follows the human-authored instructions, generates and refines execution plan by checking the case information and tools, and suggests the case resolution method autonomously at the end.