eBPF tooling like Tetragon is excellent at seeing what a workload does. For sandboxing an agent, seeing is the wrong verb. Here is why the enforcement point has to sit in the syscall path itself.
July 5, 2026
Joseph Call
Building agentic systems. Researching how they break.
Now
I build agentic AI systems for production: ingestion at scale, retrieval, and getting multiple agents to work together without falling over. I'm also a PhD researcher at William & Mary, where I work on what happens to authorization when agents start delegating to each other.
I write about both, plus travel, and whatever else I've been thinking about.
Currently: writing papers on agentic protocols over networks, building AI-powered file ingestion systems at a scale where the edge cases stop being edge cases, and spending as much time in Lisbon as I can arrange.
Selected writing
I moved to Lisbon for a while and learned the language on the way. The language was not the point. The people were.
July 22, 2025
I spent a while assuming that being good at the work was the hard part of getting the work. It is not, and figuring out why changed how I approach every conversation.
November 6, 2024
Selected work
Agents over the Network
In preparation · First author
An environment that secures agent-to-agent communication, blocking prompt injection, tool misuse, and privilege escalation.
Search-Based Evolutionary Data Pruning for Class-Level Code Summarization
ICSE 2026, JAWS Workshop · First author
Makes code-summarization models cheaper to train by finding the smallest useful training set.
Resource-Efficient and Effective Code Summarization
FORGE 2025 · Second author
Shows code summarization can be done accurately with far less compute.