AI Agents & Task Automation
Researching robust AI agent architectures that can plan, execute, and recover from multi-step tasks in desktop, web, and robotic environments.
Research Area
We investigate AI agent systems capable of reliably executing complex, multi-step tasks in open-ended environments. Our focus is on making AI agents that work in practice — not just in benchmarks.
Key Research Questions
- How can agents handle unexpected failures and recover gracefully?
- What planning representations enable efficient task decomposition?
- How do we evaluate agent reliability across diverse task domains?
- Can agents learn from user feedback to improve over time?
Current Work
Workflow Execution Engine
Building a robust execution engine for desktop and web automation tasks, including error detection and retry strategies.
Task Representation
Exploring structured task representations that balance expressiveness with executability.
Human-Agent Collaboration
Studying how users and AI agents can collaborate effectively, including clarification strategies and confidence communication.
Applications
This research directly informs the Workflow Agent product, providing the algorithmic foundation for reliable automation.