We study how agents behave when they use tools and change things around them. Where does a review help, what does it miss, and how can we measure the difference?
Our preliminary DARC study evaluates runtime controls on 34,734 action-level examples converted from four public agent-safety benchmarks.
Read the research noteThese are preliminary results on converted examples, not proof of safety in live deployments.
Our agenda covers individual tool calls and the larger systems they belong to: permissions, memory, other agents and the people responsible for the work.
How should a system review a tool call before an agent changes a file, sends a message or touches a live service?
A tool's reach depends on its permissions. We study how private context, memory and access to sensitive resources change the review.
One agent's action can become another's input. We examine delegation, shared tools and failures that spread across a workflow.
We investigate what internal signals, action history and policy matches can tell us about the risk of a proposed action.
Who created an agent? What could it access, and who approved its actions? We study how to keep those connections traceable.
Evaluations, traces and human reviews can inform deployment decisions. We examine what that evidence supports and where it falls short.
We want the research to be useful outside the paper. Our work is designed to produce better evaluations, practical review interfaces and evidence that teams can examine.
Explore our open sourceEvaluations that test what an agent does, including the tools and environment around it.
Policy and review interfaces that developers can place before execution.
Shared ways to describe risky actions and the evidence collected during a review.
Methods, results and limitations that other researchers can inspect.
If you build agent systems, design evaluations or study AI safety, we'd like to hear about your work.
