Research
Four threads, one question: what artifact would convince someone who does not trust the system, or us?
How agents behave
Agents plan, adapt, and sometimes deceive
Agents don't just follow fixed instructions. They plan, adapt, and when many of them interact they can cooperate or collude. We study what agents really do, and where that goes wrong.
Lying with Truths: Multi-Agent Collusion for Belief Manipulation (ACL 2026, Outstanding Paper)
Tapas are Free! Training-Free Adaptation of Programmatic Agents (AAAI 2026)
Building better coding agents
The agents that write the software
The most useful agents are the ones that write code, and stronger coding agents make stronger general-purpose agents. We work on how they plan, how they reuse learned skills, and how to measure whether any of it actually helps.
SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding (arXiv:2601.22956)
Skill Coverage: A Test Adequacy Metric for Agent Skills (arXiv:2606.20659)
SWE-Skills-Bench: Do Agent Skills Actually Help? (arXiv:2603.15401)
Verifying agent-written code
Code is not correct because an agent wrote it
We build tools that check generated code automatically, pairing language models with bounded model checking and symbolic execution so that the verifier, not the model, has the final word on soundness.
Agentic Model Checking (arXiv:2605.21434)
Agentic Planning for Symbolic Execution (arXiv:2608.06397)
Partial Contracts Suffice: Sound, LLM-Inferred Regression Verification (arXiv:2607.10291)