Agent Sight is a zero-instrumentation AI Agent observability tool based on eBPF, part of Alibaba's ANOLISA project. It captures LLM API calls, Token consumption, and process behavior at the kernel level without modifying Agent code, providing full-stack observability for AI Agents running on Linux.

Capabilities include multi-dimensional Token accounting by agent, task, and model; complete auditing of LLM calls and process execution; automatic discovery of running Agent processes; and interruption detection covering LLM errors, SSE truncation, context overflow, and crashes. A web Dashboard visualizes real-time Token trends, Agent health, and session traces.

Requirements include Linux kernel 5.8 or newer with BTF support, plus root or CAP_BPF privileges. Install via `sudo anolisa install agentsight`, then enable the systemd unit to run eBPF tracing and the Dashboard together. The Dashboard is served on port 7396: localhost access is authentication-free, while remote access requires an auto-generated token persisted to .dashboard_token. On macOS, only the trace and serve commands are available, without eBPF.

CLI commands such as summary, token, audit, discover, and interruption query usage and events over configurable time windows. Configuration lives in /etc/agentsight/config.json; user config files replace, not extend, built-in default rules.