Overview
My research focuses on trustworthy language-model agents. I study how agents plan, use tools, manage context, and make decisions about sensitive information in realistic settings. The aim is both empirical and practical: to identify why agents fail and to build methods that make their behavior easier to evaluate, diagnose, and improve.
Agent harnesses and evaluation
Agent performance is shaped by more than the underlying model. Planning scaffolds, tool interfaces, context-management policies, and safety checks all affect what an agent can accomplish and at what cost. In our recent study of coding-agent harnesses, we systematically vary planning, action space, and context management while holding the execution loop fixed. This work investigates how those design choices interact with model capability, task type, and context-window budget.
I am particularly interested in component-level evaluation: methods that say not only whether a system succeeds, but which intervention changed its trajectory and why.
Contextual privacy and safety
Autonomous agents increasingly operate across email, scheduling, documents, code, and other settings where information use must be appropriate to context. My earlier work at Microsoft Research Asia examined the gap between agents’ privacy judgments and their privacy actions in realistic, multi-step settings. This led to Privacy in Action, which develops realistic privacy mitigation and evaluation for LLM-powered agents.
My work on MPCI-Bench extends this question to multimodal agents. It evaluates whether agents’ information use respects the norms of the particular context in which text, images, and other signals appear—not merely whether a piece of information is sensitive in isolation.
From principles to behavior
High-level AI safety policies are necessary but insufficient on their own. A central challenge is translating principles into behavioral rules that can be tested in concrete workflows. During my internship at Zoom Video Communications, I worked on operationalizing safety policies for agentic systems and on interpreting coding-agent behavior across planning, tool-interface, and context-management settings.
Reliable agents need more than broad principles: they need behaviorally grounded evaluations that account for the task, tools, and context in which an action occurs.
Looking forward
I hope to contribute tools and empirical frameworks for building agents that are capable, transparent, and appropriately constrained. I am especially excited by work at the intersection of long-horizon agent evaluation, context-aware safety, and post-training methods for reliable behavior.
I welcome conversations and collaborations on LLM agents, agent evaluation, privacy, safety, and alignment. Please feel free to get in touch.