Agent systems
We study how models use tools to take on complete tasks and how multiple agents coordinate. CyberGuard is our open-source work in this direction, organizing security investigation, review, action proposals, and post-execution checks into a workflow.
When multiple agents take on a task, how should a system coordinate responsibilities, tool use, and follow-up actions?
As agents move from individual answers to ongoing tasks, capabilities beyond the model become essential. We explore coordination, tool integration, runtime management, and execution feedback through concrete workflows.
Make collaboration part of the workflow.
Organize roles, context, and intermediate artifacts around a task so multiple agents can keep investigations and analysis moving.
Connect tools and execution environments.
Bring model proposals into explicit tools and execution processes, including review before action and feedback afterward.
Extract infrastructure from system practice.
Use concrete projects to reveal integration and runtime challenges, then explore reusable interfaces, components, and developer experiences.
CyberGuard
Agents analyze investigation materials and produce action proposals. Approved proposals enter execution, followed by an independent check of the result. The public repository shows how this workflow is implemented and deployed.
Explore the open-source project
About the project and demo
The public repository provides project documentation, deployment materials, and a recorded demo of the workflow.