eBPF telemetry for AI agents that refuse to be instrumented

What it does

AgentSight is a system-level observer for AI agent binaries like Claude Code, Gemini CLI, or Codex. It uses eBPF and TLS traffic tracing to record process trees, file operations, network destinations, and LLM prompts/responses from outside the process, then stores everything in local SQLite databases for later inspection. It essentially treats closed-source agent CLIs as opaque black boxes and extracts their behavior at the kernel boundary.

The interesting bit

The clever part is how it bypasses the usual SDK-or-proxy trap: by hooking SSL/TLS calls at the library level, it captures plaintext LLM traffic even when agents statically link BoringSSL or OpenSSL, and it correlates those network events with subprocess spawns and file system activity in a single timeline. There is also an offline agentpprof mode that turns local agent session history into token-cost flamegraphs.

Key highlights

• Zero SDK or vendor integration required; works on unmodified, closed-source CLI agents.
• Captures kernel-level events—processes, file access, network destinations—alongside LLM prompts, tool calls, and token counts.
• Exports OpenTelemetry GenAI spans over OTLP for backend integration without in-process instrumentation.
• Includes a local web UI with timeline, process tree, and metrics views for live or recorded sessions.
• Claims less than 3% CPU overhead for typical traced workloads.

Caveats

• Linux only, and eBPF probes require root privileges (the monitored agent itself stays unprivileged).
• Statically linked SSL libraries (e.g., in Claude Code or Node.js agents) may need explicit binary path hints in attach mode, though the record command auto-discovers many cases.
• Captured data can include prompts, paths, headers, and network targets, so session databases should be treated as sensitive.

Verdict

Worth a look if you run opaque AI agent CLIs and need to audit their system behavior, token burn, or subprocess explosions. Skip it if you are on macOS or Windows, or if your agents are already fully instrumented in code you own.