AgentLens

A real-time observability and debugging layer for AI agents.

Modern agent systems are difficult to debug. Developers often cannot easily see:

• what reasoning steps an agent took
• which tools were called
• how state or memory changed
• why a decision was made

AgentLens adds a lightweight instrumentation layer that records structured events during agent execution and streams them to a dashboard.

Quick Start

Clone the repository and run the example agent to see how AgentLens captures events from tool execution.

git clone https://github.com/joshuamlamerton/AgentLens cd AgentLens python examples/simple_agent_demo.py

---

Architecture

flowchart TB

A[AI Agent]

B[AgentLens Instrumentation]

C[Event Stream]

D[Dashboard]

E[Tools / APIs]

A --> B
B --> C
C --> D

A --> E
B --> E

Loading

AgentLens sits between the agent and the tools it uses. Every action is recorded as a structured event.

---

Example Event

{
  "event": "tool_call",
  "agent": "shopping_agent",
  "tool": "search_product",
  "input": {"product": "laptop"},
  "timestamp": "2026-03-11T12:00:00Z"
}

Events like these allow developers to reconstruct what an agent did step by step.

---

Running the Demo

Clone the repository and run the example agent.

python examples/simple_agent_demo.py

Output will look similar to:

EVENT: tool_call
EVENT: tool_result
Tool result: Found laptop for $1200

This demonstrates how AgentLens records agent activity.

---

Repository Structure

agentlens

docs
  architecture.md

core
  instrumentation.py
  event_stream.py

dashboard
  server.py

examples
  simple_agent_demo.py

---

Roadmap

Phase 1
Instrumentation layer for agents

Phase 2
Local dashboard for event visualization

Phase 3
Multi-agent tracing

Phase 4
Production observability for agent systems

---

License

Apache 2.0