# AgentSight by Eunomia

AgentSight is a system-level profiling/tracing toolkit for AI agents, built from tools and skills.

Primary website: https://agentsight.us/
Live demo: https://app.agentsight.us/
Documentation: https://eunomia.dev/agentsight/
GitHub: https://github.com/eunomia-bpf/agentsight

## What AgentSight Does

- Profiles AI agent runs across time, tokens, commands, files, network calls, and system resources.
- Combines eBPF profiling tools with agent-readable skills that drive profiling workflows.
- Bridges the semantic gap between natural language prompt intent and low-level OS side effects.
- Traces Claude Code, Codex, Gemini CLI, OpenCode, OpenClaw, Python agents, Node.js agents, Dockerized agents, and arbitrary local commands.
- Uses eBPF, process monitoring, and TLS traffic tracing to observe agents from the system boundary.
- Requires Linux and elevated privileges for live eBPF capture.
- Stores recorded sessions locally and can serve or export reports from saved data.
- Exports captured LLM calls as OpenTelemetry GenAI spans.
- Supports skill-driven profiling workflows where an agent chooses tools, gathers evidence, ranks metrics, redacts sensitive data, and creates shareable HTML insight artifacts.

## Positioning

AgentSight is not an SDK-based LLM tracing platform, gateway, or transcript-only token tracker. It complements Langfuse, LangSmith, Phoenix, Helicone, Datadog, New Relic, and vendor-native telemetry by capturing system-level effects that app traces often miss: child processes, shell commands, file access, network destinations, and resource usage. AgentSight is designed for observer-agent workflows where an independent agent uses AgentSight tools and skills to connect prompt intent with raw system events.

## Common Use Cases

- Profile slow or expensive AI agent runs.
- Understand token loops, repeated file scans, and long-running subprocesses.
- Review AI-generated pull requests with a run profile.
- Generate single-file HTML insight artifacts from saved traces or monitor databases.
- Build causally linked Agent Flamegraphs from prompt intent and OS-level side effects.
- Trace closed-source local agent CLIs without vendor hooks.
- Audit MCP servers, skills, and plugins during real agent use.
- Investigate risky file, process, and network behavior.

## Short Description

Let your agents profile AI agents like programs. AgentSight provides system-level profiling/tracing tools and skills for Claude Code, Codex, and other local agents.