agentlog

OpenTelemetry-native observability for multi-agent LLM systems.

Install

pip install agentlog[openai]     # for OpenAI
pip install agentlog[litellm]    # for LiteLLM (covers Bedrock, Anthropic, Gemini)

Quick Start

import agentlog
agentlog.init()  # patches LLM clients, exports spans to collector at localhost:4317

import openai
client = openai.OpenAI()
response = client.chat.completions.create(model="gpt-4o", messages=[...])
# ^ this call is now automatically traced

Multi-Agent Tracing

@agentlog.agent(name="supervisor")
def supervisor(input):
    # all LLM calls here are grouped under "supervisor"
    ...

@agentlog.agent(name="researcher")
def researcher(query):
    # all LLM calls here are grouped under "researcher"
    ...

What Gets Captured

Every LLM call emits a span with:

• Full prompt (messages array)
• Full response (content + tool calls)
• Model name (requested and resolved)
• Token counts (input + output)
• Finish reason
• Timing and parent-child relationships

Infrastructure

The SDK sends spans to an OpenTelemetry Collector. Deploy the full stack (collector + ClickHouse + trace viewer UI) via the Helm chart:

helm install agentlog oci://ghcr.io/medhaj-ops/agentlog/chart

Or run locally with Docker Compose:

git clone https://github.com/Medhaj-ops/agentlog && cd agentlog
docker compose up -d

Links

• GitHub
• Documentation

License

Apache 2.0