Weave by Weights & Biases

Weave is a toolkit for tracing and evaluating AI agents, built by Weights & Biases.

You can use Weave to:

• Trace agents. Conversations, turns, LLM calls, and tool calls show up in the Agents tab.
• Evaluate agents and LLM applications.
• Trace functions with @weave.op when you want the Calls tab, not the Agents tab.

Documentation

Our documentation site can be found here.

Start here for agents:

• Choose an agent integration — OpenAI Agents SDK, Claude Agent SDK, and Google ADK. weave.init() is enough for those SDKs.
• Custom agents — wrap your own loop with start_conversation, start_turn, start_llm, and start_tool.

Prerequisites

• Python 3.10 or higher
• A Weights & Biases account (free tier available)
• A W&B API key from https://wandb.ai/authorize. Set WANDB_API_KEY.
• An OpenAI API key in OPENAI_API_KEY, for the example below.

Quick start: trace an agent

This example uses the OpenAI Agents SDK. The Python import is agents; the package name is openai-agents. Weave autopatches it after weave.init(). Traces land in the Agents tab, not the Calls tab.

pip install weave openai-agents requests

import asyncio
import requests
import weave
from agents import Agent, Runner, function_tool

weave.init("/")


@function_tool
def wikipedia_search(query: str) -> str:
    """Search Wikipedia for a topic and return its title and intro paragraph."""
    r = requests.get(
        "https://en.wikipedia.org/w/api.php",
        params={
            "action": "query",
            "generator": "search",
            "gsrsearch": query,
            "gsrlimit": 1,
            "prop": "extracts",
            "exintro": True,
            "explaintext": True,
            "format": "json",
        },
        headers={"User-Agent": "weave-demo"},
    ).json()
    page = next(iter(r["query"]["pages"].values()))
    return f"{page['title']}: {page['extract']}"


agent = Agent(
    name="Research assistant",
    instructions=(
        "You are a research assistant. Use the wikipedia_search tool to look up "
        "topics when needed, and cite the article titles you used."
    ),
    tools=[wikipedia_search],
)


async def main():
    history = []
    for question in [
        "Who founded Anthropic?",
        "What is Claude (the AI assistant)?",
        "Summarize what we discussed in one sentence.",
    ]:
        history.append({"role": "user", "content": question})
        print(f"USER: {question}")
        result = await Runner.run(agent, input=history)
        print(f"AGENT: {result.final_output}\n")
        history = result.to_input_list()


asyncio.run(main())

weave.init() prints a project URL. Open the Agents tab.

Claude Agent SDK works the same way: install the framework, call weave.init(), run the agent. For Google ADK, import google.adk before weave.init(). See Choose an agent integration.

Custom agents

If you are not using a supported agent SDK, wrap your own loop. This is the Conversation SDK, not the @weave.op path.

Use start_conversation, start_turn, and start_llm. In Python they are context managers and close on exceptions. start_session and Session still exist; they emit a DeprecationWarning. Do not use them in new code.

import weave

weave.init("/")

with weave.start_conversation(agent_name="research-bot"):
    with weave.start_turn(user_message="Who founded Anthropic?"):
        with weave.start_llm(model="gpt-4o-mini", provider_name="openai") as llm:
            llm.output("Anthropic was founded by former OpenAI researchers.")
            llm.record(
                usage=weave.Usage(input_tokens=12, output_tokens=9),
                response_model="gpt-4o-mini",
            )

Always pass provider_name. Weave does not infer it from the model name.

For a multi-turn loop with tools, see the custom agents quickstart.

Function tracing

@weave.op traces a function. Those traces land in the Calls tab, not the Agents tab. Use it for evaluations, scorers, and LLM calls that are not an agent loop.

Plain openai is also auto-traced into the Calls tab after weave.init(). @weave.op wraps the call in a parent function, so the OpenAI request sits under extract_fruit.

pip install weave openai

import json
import weave
from openai import OpenAI

weave.init("/")


@weave.op
def extract_fruit(sentence: str) -> dict:
    client = OpenAI()
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {
                "role": "system",
                "content": (
                    "You will be provided with unstructured data, and your task is to parse "
                    "it into one JSON object with fruit, color and flavor as keys."
                ),
            },
            {"role": "user", "content": sentence},
        ],
        temperature=0.7,
        response_format={"type": "json_object"},
    )
    extracted = response.choices[0].message.content
    return json.loads(extracted)


extract_fruit(
    "There are many fruits that were found on the recently discovered planet Goocrux. "
    "There are neoskizzles that grow there, which are purple and taste like candy."
)

Contributing

Interested in pulling back the hood or contributing? Awesome, before you dive in, here's what you need to know.

We're in the process of 🧹 cleaning up 🧹. This codebase contains a large amount code for the "Weave engine" and "Weave boards", which we've put on pause as we focus on Tracing and Evaluations.

The Weave Tracing code is mostly in: weave/trace and weave/trace_server.

The Weave Evaluations code is mostly in weave/flow.