Inspiration
Email and google drive are already trusted by the UCSC campus community for moving and storing important data, but email is distracting and overwhelming, and relies on an actual human on the other end to respond and is slow and pretty delayed. We want agents to be able to push information to one another so that it is ready at hand only when you need it.
Tarun is a student and Adam is a teacher, so the two of us on the team could imagine may scenarios where we might want to help a student’s agentic knowledge base stay up to date without human attention every time. We wanted something like email but fit for the emerging agent-literate UCSC community.
What it does
Campus Cubbies is an extension to the (existing) BayLeaf AI platform to let agents share information with one another. Agents discover our system's capabilities automatically through A2A, figuring out how to deposit information into other user’s agent cubbies and controlling who else can contribute to their own. Users are identified by there existing @ucsc.edu Google accounts.
We imagine it could be used by international students (like Tarun) to receive pushed updates fro the International Students Office (ISO), but that’s just one of a zillion possible uses for identity-verified peer-to-peer communication between agents on campus.
How we built it
Hosted infrastructure: Google Cloud Run (hosts the server side of our A2A endpoints) Google Drive (stores user cubby data in a form that is easy to inspect with high availability and approval by campus IT to store student data) Google Firestore (stores cubby metadata like who is allowed to contribute who who else’s cubby)
Coding agents: Claude Code + Anthropic API / OpenChamber + BayLeaf API
Runtime inference: Entirely provided by end users who connect with their already-capable agents.
Challenges we ran into
The sandboxes provided by many retail generative AI systems don’t allow general outbound network connections, effectively blocking A2A and other protocols. But Bayleaf Sandboxes allow general network egress!
Many agents stumbled over tools during the onboarding flow or failed to give users a clean narration of what was happening. But a few revisions of our skill text fixed that, and it encouraged the user’s agent to write a local skill to speed up the process in the future.
Accomplishments that we're proud of
Integration with actual UCSC systems: Single Sign-On and Google Drive.
Verified compatibility with many agent harnesses.
What we learned
Compelling agent-to-agent applications should be derived from real world use cases, not imaginary scenarios. But now we have a real system we can use for actual campus business to see what sort of interesting trouble we get into!
What's next for GremlinLabs - Campus Cubbies
We’re going to leave the service up and running, already fully integrated with UCSC IT systems.
Log in or sign up for Devpost to join the conversation.