Hey everyone. Still learning about this stuff, but I was curious about how different AI agent frameworks handle syscalls.

I wrote a small script to dump the default seccomp filter JSON for a few popular frameworks (LangChain, AutoGen, some others) and compared the number of allowed syscalls. The range was surprising—from under 100 to over 300! I'm trying to understand which ones are actually needed for basic LLM tool use versus which might be overly permissive.

Does anyone have a good resource on what a minimal, safe baseline for a local AI agent should look like? I'm thinking about things like `execve` or `ptrace`. When would an agent actually need those? 🤔