agentsight vs Tracekit
System-level observability for AI agents via eBPF and TLS tracing: correlates prompts and model calls with the processes, files and network the agent actually touched — no SDK, no proxy. — versus — Flight recorder for AI agents: policy-gates every tool call before it runs, holds risky ones for human approval, and signs each action into a tamper-evident log anyone can verify offline.
Both answer 'what did my agent actually do'; AgentSight observes from outside via eBPF, Tracekit sits in the tool path and can block and sign.
| agentsight | Tracekit | |
|---|---|---|
| Stars | 798 | 734 |
| Forks | 121 | 130 |
| Language | C | Python |
| License | MIT | MIT |
| Last activity | today | today |
| Topics | security, coding, local | security, evals |
| Curated connections | 9 | 5 |
agentsight — the curator's take
Use it when the agent is a closed CLI (Claude Code, Codex, Gemini CLI, OpenClaw) and its own logs don't explain a failure: AgentSight watches at the kernel boundary, so it sees the model traffic *and* every exec, file write and outbound connection, and ties them together. That's the layer LangSmith-class tools can't reach without owning the app code. NOT for prompt/eval-level tracing of an app you wrote (use an OTLP tracer), and it's Linux-with-eBPF: on macOS you're out of luck. Not an enforcement tool — it observes and advises, it doesn't block.
Tracekit — the curator's take
Use it when agents touch real systems and you must prove to customers or auditors what happened — the signer sits outside the agent's reach, so a compromised agent can't rewrite history. `tracekit audit --since 7d` over your Claude Code logs is a zero-install way to see what it would have caught. Overkill for local tinkering; it gates and records, it doesn't sandbox — pair with real isolation for untrusted code.