145b1669e7
Adds metadataBase, full OpenGraph + Twitter card tags, keywords, JSON-LD structured data (SoftwareApplication + Organization), sitemap.ts, robots.ts with AI crawler directives, and llms.txt for AI agent discoverability.
51 lines
3.0 KiB
Plaintext
51 lines
3.0 KiB
Plaintext
# AgentLens
|
|
|
|
> AgentLens is an open-source agent observability platform that traces AI agent decisions, not just API calls. It captures why agents choose specific tools, routes, or strategies — providing visibility into the reasoning behind every action.
|
|
|
|
AgentLens helps engineering teams debug, monitor, and improve AI agent applications in production. Unlike traditional LLM observability tools that only trace API calls, AgentLens captures the decision-making process: tool selection rationale, routing logic, retry strategies, and planning steps. It includes a real-time dashboard with decision tree visualization, cost analytics, and token tracking.
|
|
|
|
## Getting Started
|
|
|
|
- [GitHub Repository](https://gitea.repi.fun/repi/agentlens): Source code, issues, and contribution guide
|
|
- [PyPI Package](https://pypi.org/project/vectry-agentlens/): Install with `pip install vectry-agentlens`
|
|
- [Dashboard](https://agentlens.vectry.tech/dashboard): Live demo dashboard with sample traces
|
|
|
|
## Python SDK
|
|
|
|
- [Basic Usage](https://gitea.repi.fun/repi/agentlens/src/branch/main/examples/basic_agent.py): Minimal SDK usage with trace context and decision logging
|
|
- [OpenAI Integration](https://gitea.repi.fun/repi/agentlens/src/branch/main/examples/openai_agent.py): Wrap OpenAI client for automatic LLM call tracing
|
|
- [Multi-Agent Example](https://gitea.repi.fun/repi/agentlens/src/branch/main/examples/multi_agent.py): Nested multi-agent workflow tracing
|
|
- [Function Calling](https://gitea.repi.fun/repi/agentlens/src/branch/main/examples/moonshot_real_test.py): Real LLM test with tool/function calling
|
|
|
|
## Key Concepts
|
|
|
|
- **Traces**: Top-level containers for agent execution sessions, with tags and metadata
|
|
- **Spans**: Individual operations within a trace (LLM calls, tool calls, chain steps)
|
|
- **Decision Points**: The core differentiator — captures what was chosen, what alternatives existed, and why
|
|
- **Decision Types**: TOOL_SELECTION, ROUTING, RETRY, ESCALATION, MEMORY_RETRIEVAL, PLANNING, CUSTOM
|
|
|
|
## API
|
|
|
|
- POST /api/traces: Batch ingest traces from SDK (Bearer token auth)
|
|
- GET /api/traces: List traces with pagination, search, filters, and sorting
|
|
- GET /api/traces/:id: Get single trace with all spans, decisions, and events
|
|
- GET /api/traces/stream: Server-Sent Events for real-time trace updates
|
|
- GET /api/health: Health check endpoint
|
|
|
|
## Integrations
|
|
|
|
- **OpenAI**: `wrap_openai(client)` auto-instruments all chat completions, streaming, and tool calls
|
|
- **LangChain**: `AgentLensCallbackHandler` captures chains, agents, tools, and LLM calls
|
|
- **Any Python Code**: `@trace` decorator and `log_decision()` for custom instrumentation
|
|
|
|
## Self-Hosting
|
|
|
|
- Docker Compose deployment with PostgreSQL and Redis
|
|
- Single `docker compose up -d` to run
|
|
- Environment variables: DATABASE_URL, REDIS_URL, AGENTLENS_API_KEY
|
|
|
|
## Optional
|
|
|
|
- [Company Website](https://vectry.tech): Built by Vectry, an engineering-first AI consultancy
|
|
- [CodeBoard](https://codeboard.vectry.tech): Sister product — understand any codebase in 5 minutes
|