dunetrace 0.2.0

Behavioral observability for AI agents

Dunetrace SDK

Behavioral observability for AI agents. Zero-dependency Python SDK that detects failure patterns in your LLM agents in real-time.

Install

pip install dunetrace

With integrations:

pip install dunetrace[langchain]   # LangChain

Quickstart

from dunetrace import Dunetrace

dt = Dunetrace(endpoint="http://localhost:8001")

with dt.run("my-agent", user_input="What is the capital of France?") as run:
    # your agent logic here
    run.emit("tool.called", {"tool_name": "web_search"})

What it detects

• Tool loops — same tool called repeatedly without progress
• Context bloat — prompt tokens growing 3× across steps
• Tool avoidance — agent has tools but never uses them
• Step count inflation — far more steps than needed
• First step failure — agent fails immediately
• RAG empty retrieval — retrieval returns no results
• And more (tool thrashing, retry storms, cascading failures)

Self-hosted backend

Dunetrace requires the backend to collect and analyze events:

git clone https://github.com/dunetrace/dunetrace
cd dunetrace
cp .env.example .env
docker compose up

Links

• GitHub
• Issues