Long-horizon coding agents
Keep extended coding tasks on course.
How can a coding agent preserve security intent, recover from mistakes and demonstrate completion across a long task?
Consider an agent updating a shared authentication mechanism across several services. It must plan the work, retain relevant context, make compatible changes and verify that security properties still hold. This exploratory area examines how agents manage such extended tasks, including intermediate checks, recovery after failed steps and decisions that require a human reviewer.
Research areas
Represent task goals, security constraints and repository knowledge so that they remain usable as the working context changes.
Break extended work into reviewable steps with checkpoints that verify progress against the original goal.
Detect accumulated errors, recover from failed tools or tests and resume without losing important evidence.
Study tool permissions, approval boundaries and human handoffs during security-sensitive code changes.
Evaluate end-to-end completion through security checks, regression tests and reproducible evidence of the changes made.
Publications & related resources
Projects, publications and research materials will be added here.
Contribute to this research.
Explore student research, internships and collaboration with the lab.