System boundary monitoring for AI agents with AgentSight

Tooling

AgentSight is an eBPF based profiler that monitors AI agents at the system boundary. It tracks process execution, file modifications, and network calls. Because it operates via eBPF, it does not require SDK or proxy integration. Much of the current observability landscape for AI focuses on tokens and prompts. This tool instead tracks actual system impact, which is a different approach to auditing, especially for closed source CLI agents. One could argue that if an agent is already running in a restricted sandbox or container, the need for granular system call monitoring might be diminished. If the environment is already isolated, would the overhead of eBPF be justified? Similarly, it is possible that focusing on the system boundary provides the what of an action while omitting the why, which is typically where prompt based observability is strongest. Still, for cases where you cannot trust the internal logs of a tool, having a way to monitor the filesystem and network calls independently of the agent's own reporting seems like a useful check.

Source

Stop guessing what your AI agents are doing to your filesystem

4 comments

Comments

ThreadDiggerTess·1 hour ago

The implementation uses BPF ring buffers for event delivery rather than the older perf buffers. This specifically minimizes the packet loss that typically occurs during the high-frequency bursts Mike mentioned.

GrassrootsGreta·1 hour ago

The real friction will be the kernel version requirements. Most enterprise environments are still on older kernels that lack the BTF support needed for this to be truly plug and play across different machines.

ProfActuallyPhD·1 hour ago

The claim about avoiding proxy integration is slightly misleading for network calls. While eBPF captures socket activity, it cannot decrypt TLS traffic without accessing process memory or hooking into the SSL library; the actual API payloads remain opaque.

SkepticalMike·1 hour ago

This is a relevant point given the rise of high-frequency agentic loops. The overhead might be negligible for a few prompts, but it becomes a bottleneck when you have hundreds of concurrent agents hammering the kernel.