AgentSight is a observability tool designed specifically for monitoring LLM agent behavior through SSL/TLS traffic interception and process monitoring. Unlike traditional application-level instrumentation, AgentSight observes at the system boundary using eBPF technology, providing comprehensive insights into AI agent interactions with minimal performance overhead.
✨ Zero Instrumentation Required - No code changes, no new dependencies, no SDKs. Works with any AI framework or application out of the box.
wget https://github.com/eunomia-bpf/agentsight/releases/latest/download/agentsight && chmod +x agentsight
# Record agent behavior from claude
sudo ./agentsight record -c "claude"
# Record agent behavior from gemini-cli (comm is "node")
sudo ./agentsight record -c "node"
# For Python AI tools
sudo ./agentsight record -c "python"
# Record claude or gemini activity with NVM Node.js, if bundle OpenSSL statically
sudo ./agentsight record --binary-path /usr/bin/node -c nodeVisit http://127.0.0.1:7395 to view the recorded data.
Real-time process tree visualization showing AI agent interactions and file operations
Real-time timeline visualization showing AI agent interactions and system calls
Real-time metrics visualization showing AI agent memory and CPU usage
| Challenge | Application-Level Tools | AgentSight Solution |
|---|---|---|
| Framework Adoption | ❌ New SDK/proxy for each framework | ✅ Drop-in daemon, no code changes |
| Closed-Source Tools | ❌ Limited visibility into operations | ✅ Complete visibility into prompts & behaviors |
| Dynamic Agent Behavior | ❌ Logs can be silenced or manipulated | ✅ Kernel-level hooks for reliable monitoring |
| Encrypted Traffic | ❌ Only sees wrapper outputs | ✅ Captures real unencrypted requests/responses |
| System Interactions | ❌ Misses subprocess executions | ✅ Tracks all process behaviors & file operations |
| Multi-Agent Systems | ❌ Isolated per-process tracing | ✅ Global correlation and analysis |
AgentSight captures critical interactions that application-level tools miss:
- Subprocess executions that bypass instrumentation
- Raw encrypted payloads before agent processing
- File operations and system resource access
- Cross-agent communications and coordination
┌─────────────────────────────────────────────────┐
│ AI Agent Runtime │
│ ┌─────────────────────────────────────────┐ │
│ │ Application-Level Observability │ │
│ │ (LangSmith, Helicone, Langfuse, etc.) │ │
│ │ 🔴 Can be bypassed │ │
│ └─────────────────────────────────────────┘ │
│ ↕ (Can be bypassed) │
├─────────────────────────────────────────────────┤ ← System Boundary
│ 🟢 AgentSight eBPF Monitoring (Kernel-level) │
│ ┌─────────────────┐ ┌─────────────────────┐ │
│ │ SSL Traffic │ │ Process Events │ │
│ │ Monitoring │ │ Monitoring │ │
│ └─────────────────┘ └─────────────────────┘ │
└─────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────┐
│ Rust Streaming Analysis Framework │
│ ┌─────────────┐ ┌──────────────┐ ┌────────┐ │
│ │ Runners │ │ Analyzers │ │ Output │ │
│ │ (Collectors)│ │ (Processors) │ │ │ │
│ └─────────────┘ └──────────────┘ └────────┘ │
└─────────────────────────────────────────────────┘
↓
┌─────────────────────────────────────────────────┐
│ Frontend Visualization │
│ Timeline • Process Tree • Event Logs │
└─────────────────────────────────────────────────┘
-
eBPF Data Collection (Kernel Space)
- SSL Monitor: Intercepts SSL/TLS read/write operations via uprobe hooks
- Process Monitor: Tracks process lifecycle and file operations via tracepoints
- <3% Performance Overhead: Operates below application layer with minimal impact
-
Rust Streaming Framework (User Space)
- Runners: Execute eBPF programs and stream JSON events (SSL, Process, Agent, Combined)
- Analyzers: Pluggable processors for HTTP parsing, chunk merging, filtering, logging
- Event System: Standardized event format with rich metadata and JSON payloads
-
Frontend Visualization (React/TypeScript)
- Interactive timeline, process tree, and log views
- Real-time data streaming and analysis
- See "Web Interface Access" section for details
eBPF Programs → JSON Events → Runners → Analyzer Chain → Frontend/Storage/Output
- Linux kernel: 4.1+ with eBPF support (5.0+ recommended)
- Root privileges: Required for eBPF program loading
- Rust toolchain: 1.88.0+ (for building collector)
- Node.js: 18+ (for frontend development)
- Build tools: clang, llvm, libelf-dev
AgentSight runs in Docker with --privileged for eBPF, --pid=host to access host processes, -v /sys:/sys:ro for process monitoring, and -v /usr:/usr:ro -v /lib:/lib:ro for SSL library access (required to attach uprobes to shared libraries like libssl.so). Example:
docker run --privileged --pid=host --network=host \
-v /sys:/sys:ro -v /usr:/usr:ro -v /lib:/lib:ro \
-v $(pwd)/logs:/logs \
ghcr.io/eunomia-bpf/agentsight:latest \
record --comm claude --log-file /logs/record.log# Clone repository with submodules
git clone https://github.com/eunomia-bpf/agentsight.git --recursive
cd agentsight
# Install system dependencies (Ubuntu/Debian)
make install
# Build all components (frontend, eBPF, and Rust)
make build
# Or build individually:
# make build-frontend # Build frontend assets
# make build-bpf # Build eBPF programs
# make build-rust # Build Rust collector
# Combined SSL and process monitoring with web interface
sudo ./agentsight trace --ssl --process --server
# Custom port and log file
sudo ./agentsight record -c "python" --server-port 8080 --log-file /tmp/agent.logFor Node.js installed via NVM, use the --binary-path option:
# Monitor Node.js applications with statically-linked SSL
sudo ./agentsight ssl --binary-path ~/.nvm/versions/node/v20.0.0/bin/node --comm node
# Record with custom binary path
sudo ./agentsight record -c node -- --binary-path ~/.nvm/versions/node/v20.0.0/bin/node# Run eBPF programs directly for development/testing
sudo ./bpf/sslsniff --binary-path ~/.nvm/versions/node/v20.0.0/bin/node --verbose
sudo ./bpf/process -c pythonAll monitoring commands with --server flag provide web visualization at:
- Timeline View: http://127.0.0.1:7395/timeline
- Process Tree: http://127.0.0.1:7395/tree
- Raw Logs: http://127.0.0.1:7395/logs
Q: How does AgentSight differ from traditional APM tools? A: AgentSight operates at the kernel level using eBPF, providing system-level monitoring that is independent of application code. Traditional APM requires instrumentation that can be modified or disabled.
Q: What's the performance impact? A: Less than 3% CPU overhead due to optimized eBPF kernel-space data collection.
Q: Can agents detect they're being monitored?
A: Detection is extremely difficult since monitoring occurs at the kernel level without code modification.
Q: Which Linux distributions are supported? A: Any distribution with kernel 4.1+ (5.0+ recommended). Tested on Ubuntu 20.04+, CentOS 8+, RHEL 8+.
Q: Can I monitor multiple agents simultaneously?
A: Yes, use combined monitoring modes for concurrent multi-agent observation with correlation.
Q: How do I filter sensitive data?
A: Built-in analyzers can remove authentication headers and filter specific content patterns.
Q: Why doesn't AgentSight capture traffic from my NVM Node.js application?
A: NVM Node.js binaries statically link OpenSSL instead of using system libraries. See the "NVM Node.js Applications" section for --binary-path usage examples.
Q: "Permission denied" errors
A: Ensure you're running with sudo or have CAP_BPF and CAP_SYS_ADMIN capabilities.
Q: "Failed to load eBPF program" errors A: Verify kernel version meets requirements (see Prerequisites). Update vmlinux.h for your architecture if needed.
We welcome contributions! After cloning and building (see Installation above), you can:
# Run tests
make test
# Frontend development server
cd frontend && npm run dev
# Build debug versions with AddressSanitizer
make -C bpf debug- CLAUDE.md - Project guidelines and architecture
- collector/DESIGN.md - Framework design details
- docs/why.md - Problem analysis and motivation
MIT License - see LICENSE for details.
💡 The Future of AI Observability: As AI agents become more autonomous and capable of self-modification, traditional observability approaches become insufficient. AgentSight provides independent, system-level monitoring for safe AI deployment at scale.