Manulife Financial to contribute $85k to enhance the robustness and real-world applicability of the agentic system
Professor Pascal Poupart has been awarded $170,000 through the NSERC Alliance Grant program. This funding is complemented by $85,000 from industry partner Manulife Financial, bringing the total project support to $255,000. Manulife is also providing significant in-kind contributions, including synthetic data, technical infrastructure for simulated business environments, and access to its technical experts.
Titled “Agentic Workflow Adaptation and Validation,” the two-year project will advance two research directions related to agentic ecosystems. First, it will develop continual learning methods that allow agents to adapt to evolving customer trends, changes to tool and large language models, and updates to application programming interfaces. Second, it will develop verification and validation methods to ensure that the actions taken by an ecosystem of agents meet desirable criteria and successfully accomplish target tasks.
The research will help Manulife automate knowledge-intensive workflows in insurance underwriting, freeing office professionals to focus on more creative aspects of their work. It will also position Canada at the forefront of agentic ecosystems, an emerging paradigm that leverages large language models to automate digital workflows.
Pascal Poupart is a Professor at the Cheriton School of Computer Science, a Canada CIFAR AI Chair at the Vector Institute, the research director of the Vector Institute, and a member of the Waterloo Data and AI Institute. His research focuses on developing algorithms for machine learning with application to natural language processing and material design. He is best known for his contributions to the development of reinforcement learning algorithms.
His notable projects include Bayesian federated learning, probabilistic deep learning, data-efficient reinforcement learning, inverse constraint learning, reward-guided text generation, agentic workflow adaptation and validation, multi-agent coordination, multi-agent LLM orchestration safety and dynamic composition of image generation models.
Context
The rise of large language models has accelerated development of agentic frameworks capable of automating workflows in the service industry. An agent can be a large language model, a software tool, or a human collaborator. Because complex tasks are often best addressed by breaking them down into smaller and simpler ones, it is desirable to use multiple specialized agents and to orchestrate their activities through coordinated interactions, often mediated by natural language communication.
To support this approach, many platforms have been designed to facilitate development and orchestration of agents for workflow automation. While these platforms allow engineers to develop specific workflows, track performance and debug agents, significant challenges remain. In particular, there is a need to ensure these workflows adapt continually to changes and that the output of individual agents is accurate, reliable and compliant with regulatory requirements.
The goal of this project is to develop a novel plug-and-play agentic framework with the following two objectives: (1) automated verification and correction of agent outputs to ensure reliability and compliance; and (2) continual adaptation that handles addition or removal of agents, application programming interface changes, tool and large language model updates, market shifts, and evolving customer needs.
Expected outcomes include development of a new agentic platform that facilitates continual adaptation and verification, demonstration of this platform in an insurance underwriting use case, public release of the agentic platform, and publications in top-tier AI venues of the algorithmic solutions. The expected impact of this work will increase adoption of agentic systems for workflow automation across the service sector. Workflow automation is expected to transform office work by enabling productivity gains in knowledge-intensive processes. Similar to how the Industrial Revolution automated labour, agentic systems will advance the automation of knowledge-intensive workflows, leading to reduced costs, increased efficiency and improved services.
The project will also contribute to the training of the next generation of researchers by supporting a master’s student, a doctoral student, and a postdoctoral researcher, strengthening Canada’s talent pipeline in artificial intelligence and agentic systems. Ultimately, it will help boost the nation’s productivity and strengthen its position as a global leader in AI-driven workflow automation.