Weave

Weave is a toolkit from Weights & Biases for tracing and evaluating AI agents and LLM applications. It ships Python and TypeScript libraries, and the Python library is released under the Apache 2.0 license. Weights & Biases is now part of CoreWeave Forge, and the documentation is hosted on the CoreWeave docs site.

What It Is

Weave is an observability and evaluation platform for agents and LLM apps. After installing the library and calling weave.init() with a project name, it records what your application does and sends the data to your W&B (CoreWeave Forge) account for review.

Tracing Agents and Functions

For supported agent SDKs, including the OpenAI Agents SDK, Claude Agent SDK and Google ADK, weave.init() is enough to trace conversations, turns, LLM calls and tool calls. These appear in the Agents tab. Custom agent loops can be wrapped with start_conversation, start_turn, start_llm and start_tool. Agents instrumented with OpenTelemetry can send spans to Weave's OTLP endpoint without installing the Weave SDK.

For non-agent code, the @weave.op decorator traces functions into the Calls tab. This suits evaluations, scorers and standalone LLM calls. Plain openai calls are also auto-traced after initialization.

Evaluation and Iteration

The docs describe evaluation pipelines that use LLM judges and custom scorers. They include tutorials for evaluating RAG applications. The platform also covers versioning, prompt and model experimentation, and monitoring with feedback collection.

Setup Path

Install with pip install weave or npm install weave, create an account, and set an API key. The Python library requires Python 3.10 or higher. The docs note that CoreWeave Agent Lens is the successor to Weights & Biases Weave for agent tracing workflows. Both services already receive agent tracing data, so nothing needs to be migrated.