Weave is a Weights & Biases toolkit for tracing and evaluating AI agents and LLM applications. Agent traces record conversations, turns, LLM calls, and tool calls, and they appear in an Agents view separate from the function-level Calls view.
For the OpenAI Agents SDK, Claude Agent SDK, and Google ADK, a single initialization call patches the framework and traces runs with no further code. A custom agent loop is traced through the Conversation SDK, which wraps conversations, turns, LLM calls, and tool calls in context managers that close on exceptions. Each LLM call records its model, provider, and token usage.
The @weave.op decorator traces any Python function into the Calls view, for evaluations, scorers, and LLM calls outside an agent loop. Plain OpenAI client calls are traced automatically after initialization. Traces capture code versions, parameters such as temperature and system prompts, inputs, outputs, and multi-turn messages.
Features
- Agent tracing: conversations, turns, LLM calls, and tool calls in a dedicated Agents view
- Framework integrations: OpenAI Agents SDK, Claude Agent SDK, and Google ADK traced after one initialization call
- Conversation SDK: context managers for tracing a custom agent loop, with token usage per LLM call
- Function tracing: a decorator that traces any Python function, including evaluations and scorers
- Automatic OpenAI tracing: OpenAI client calls are traced without extra code
- Evaluations: evaluate agents and LLM applications with scorers, including LLM-as-judge scoring
