AgentSight Brings Kernel-Level AI Agent Tracing Without Touching Your Code

3. AgentSight Brings Kernel-Level AI Agent Tracing Without Touching Your Code

Alibaba's open-source project Anolisa has shipped AgentSight, a zero-instrumentation observability tool for AI agents that operates at the Linux kernel level using eBPF. It captures LLM API calls, token consumption by agent, task, and model, plus full process execution traces without any modification to agent code. It requires Linux kernel 5.8 or higher with BTF support, runs as a systemd service on port 7396, and ships a web dashboard for real-time token trends and session traces. The GitHub repository sits at 375 stars and 95 forks. HN engagement is modest at 14 points, but the technical surface area is real.

The strategic edge here is architectural independence. Today's dominant observability approaches for AI agents, including LangSmith from LangChain, Arize Phoenix, and Weights & Biases Weave, all require SDK-level instrumentation. That means teams must wrap their LLM calls, modify their agent harness, and accept a dependency on the observability vendor's abstractions. AgentSight bypasses that entirely by intercepting at the kernel level. For teams running agents inside containers or on shared infrastructure where modifying the agent codebase is operationally expensive or politically blocked, that is a meaningful shift in who controls the tracing layer. Alibaba is effectively proposing that observability should be an infrastructure concern, not a framework concern.

The broader pattern worth watching: eBPF has already eaten network observability (Cilium, Falco, Pixie) and is now moving into application-layer tracing. Applying it to LLM agent workloads is a logical next step, and AgentSight's detection of SSE truncation, context overflow, and LLM errors at the kernel level suggests the approach can surface failure modes that SDK-level tools miss entirely. If the Linux-only constraint loosens and macOS eBPF support matures, the addressable deployment surface grows substantially.

Source: AgentSight - eBPF observability for AI agents (alibaba/anolisa)