HTTP 200 does not mean the agent was right.
Most agent errors look successful at the API layer. The bugs happen inside the decision chain.
Observability for agents
Standard observability doesn't work for agents.
Traditional APM tools track network requests and system metrics. They can't show you why an agent made a specific decision, which tool was invoked with what arguments, or how a prompt led to a failure.
Agents fail across multiple steps.
A single query triggers a chain of LLM calls, tool executions, and vector DB queries. Traditional logs miss the context flow.
Debugging agents is still too manual.
Developers spend hours stitching raw logs, database states, prompt strings, and JSON traces together after a run fails.
What AgentLens does
One trace layer across every agent run.
AgentLens sits beside your agent framework and automatically captures the execution graph of your workflows. It is built for developers who need to debug fast and understand exactly how their agents behave.
OpenAI
Anthropic
Gemini
LangGraph
CrewAI
FastAPI
PostgreSQL
agent-session-8fa21 / trace
Prompt captured
System, user, retrieved context, and model parameters preserved.
Tool chain mapped
Every API call, function argument, input, and output linked to the step.
Cost calculated
Token consumption, model pricing, and execution latency tracked in real time.
Error tracked
Exceptions, rate limits, and API failures logged with full stack context.
Replay engine
Reconstruct a run exactly as it happened, step by step, without asking engineers to guess the agent's internal state.
Span trees
Visualise nested loops, parallel tool calls, and complex agent reasoning paths in a developer-friendly UI.
Zero-config setup
Instrument your OpenAI, Anthropic, or custom client with just a few lines of code and immediately see results.
Who needs this first
Built for teams building production-ready AI agents.
AI platform teams
Standardize tracing, replay, and debugging across all your internal agent applications.
Software engineers
Debug multi-agent reasoning paths, parallel tool executions, and state variables in a unified trace UI.
Founders and builders
Ship agentic workflows faster with the observability layer you need to scale from MVP to production.
Product managers
Analyze token consumption, evaluate agent success rates, and monitor real-user interactions.
The thesis
The hardest part of building AI agents isn't the model. It's understanding what they do.
Start tracing in minutes
Understand your agentic workflows today.
Install the open-source SDK, initialize it in your code, and see every prompt, tool call, and token cost visualised instantly.
quickstart.py
# 1. Install agentlens
$ pip install agentlens
# 2. Add to your script
import agentlens as al
al.init(api_key="al_live_...")
al.instrument_openai() # traces all calls