Management & Observability SDK for AI Agents in production
The AgentBasis Python SDK provides a simple, lightweight way to track the performance, traces, sessions, and behavior of AI agents. It sends data using the OpenTelemetry (OTel) standard, making it compatible with AgentBasis and other observability backends.
This is the foundation SDK that enables deep observability for coded agents built with:
- Pure Python
- LLM Providers:
- OpenAI
- Anthropic
- Gemini
- Frameworks
- LangChain
- Pydantic AI
pip install agentbasisStart by initializing the SDK with your API key and Agent ID. This usually goes at the top of your main application file.
import agentbasis
# Initialize with your API Key and Agent ID
agentbasis.init(
api_key="your-api-key-here",
agent_id="your-agent-id-here"
)Use the @trace decorator to automatically track any function.
from agentbasis import trace
@trace
def chat_with_user(query):
# Your agent logic here
return "Response to: " + query
# When you call this, data is automatically sent to AgentBasis
chat_with_user("Hello world")Automatically track all your OpenAI calls (models, tokens, prompts) with one line of code.
from agentbasis.llms.openai import instrument
# Enable OpenAI instrumentation
instrument()
# Now just use the OpenAI client as normal
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4",
messages=[{"role": "user", "content": "Hello"}]
)Automatically track all your Anthropic Claude calls.
from agentbasis.llms.anthropic import instrument
# Enable Anthropic instrumentation
instrument()
# Now just use the Anthropic client as normal
from anthropic import Anthropic
client = Anthropic()
response = client.messages.create(
model="claude-3-opus-20240229",
max_tokens=1024,
messages=[{"role": "user", "content": "Hello"}]
)Track chains, tools, retrievers, and LLM calls in LangChain with full parent-child span relationships.
from agentbasis.frameworks.langchain import get_callback_handler
# Create a callback handler
handler = get_callback_handler()
# Pass it to your LangChain calls
from langchain_openai import ChatOpenAI
llm = ChatOpenAI(model="gpt-4")
response = llm.invoke("Hello world", config={"callbacks": [handler]})For chains and agents, pass the callback handler in the config:
from langchain.chains import LLMChain
from langchain.prompts import PromptTemplate
from agentbasis.frameworks.langchain import get_callback_config
# Use get_callback_config() for convenience
chain = LLMChain(llm=llm, prompt=PromptTemplate.from_template("{query}"))
result = chain.invoke({"query": "What is AI?"}, config=get_callback_config())Track Pydantic AI agents with built-in OpenTelemetry support.
from agentbasis.frameworks.pydanticai import instrument
# Enable global instrumentation for all Pydantic AI agents
instrument()
# Your agents are now automatically traced
from pydantic_ai import Agent
agent = Agent("openai:gpt-4")
result = agent.run_sync("Hello!")For per-agent control with user context:
from agentbasis.frameworks.pydanticai import create_traced_agent
# Create an agent pre-configured with tracing and context
agent = create_traced_agent(
"openai:gpt-4",
system_prompt="You are a helpful assistant."
)
# Set user context - it will be included in traces
agentbasis.set_user("user-123")
result = agent.run_sync("Hello!")Associate traces with specific users and sessions to debug issues and see per-user analytics.
# Set the current user (from your auth system)
agentbasis.set_user(current_user.id)
# Optionally set session and conversation IDs
agentbasis.set_session("session-abc")
agentbasis.set_conversation("conv-123")
# All subsequent LLM calls will be tagged with this context
response = client.chat.completions.create(...)Or use the context manager for scoped context:
from agentbasis import context
with context(user_id="user-123", session_id="session-abc"):
# All traces in this block include the context
response = client.chat.completions.create(...)- OpenTelemetry: We use OTel under the hood for maximum compatibility.
- Spans: Every action (function call, LLM request) is recorded as a Span.
- Transport: Data is batched and sent asynchronously to the AgentBasis backend.
For full documentation, visit docs.agentbasis.co.
See CONTRIBUTING.md for development setup and guidelines.
MIT License - see LICENSE for details.