wandb/weave

Weave by Weights & Biases

What it is – Weave is a Python toolkit that records ("traces") the behaviour of AI agents and the functions they call. The recorded data appears in the W&B UI under Agents (for full conversational runs) or Calls (for individual function calls), letting you inspect, debug, and evaluate your agents.

Key capabilities

• Agent tracing – automatically logs each turn, LLM request, and tool use when you run an agent built with the OpenAI Agents SDK, Claude Agent SDK, or Google ADK. The UI shows a timeline of the conversation.
• Custom‑agent support – if you have your own loop, you can wrap it with weave.start_conversation, weave.start_turn, and weave.start_llm (context‑manager style) to get the same visualisation.
• Function tracing – decorate any Python function with @weave.op to capture its inputs, outputs, and any nested LLM calls. These appear in the Calls tab and are useful for evaluation or scoring pipelines.
• Built‑in evaluation – the weave/flow package provides utilities for scoring agent outputs, though the README only hints at this.
• Zero‑config for supported SDKs – a single weave.init("<team>/<project>") call patches the supported agent libraries, so no code changes are required beyond the init.

Typical workflow

1. Install the library (and the agent SDK you need). ```bash pip install weave openai-agents # or claude‑agents, google‑adk

2. Call weave.init("myteam/myproject") – this prints a URL to the W&B project.
3. Run your agent as usual. Weave automatically records the conversation.
4. Open the project URL and explore the Agents tab to see a step‑by‑step view of prompts, tool calls, and LLM usage.
5. For non‑agent code, wrap functions with @weave.op or use the low‑level weave.start_* context managers.

Example – tracing a simple research‑assistant agent that can call a Wikipedia search tool. The README shows a complete script that:

• defines a wikipedia_search tool,
• creates an Agent with instructions,
• runs three questions in a loop using Runner.run,
• and prints the agent’s responses. After weave.init the whole interaction appears in the W&B UI.

Who it’s for

• ML engineers building LLM‑powered assistants or chatbots who need visibility into prompt flow and tool usage.
• Researchers evaluating different prompting strategies or model versions.
• Teams that already use Weights & Biases for experiment tracking and want a unified place for agent observability.

Prerequisites

• Python 3.10+.
• A free W&B account and API key.
• API keys for the LLM provider you plan to use (e.g., OpenAI).

Documentation & support

• Full docs: https://wandb.me/weave
• Quick‑start guides for each supported agent SDK.
• Open‑source repository on GitHub (wandb/weave) with CI badges and a PyPI release.

Status – The README notes that the broader “Weave engine” and UI boards are on pause; the current focus is on the tracing and evaluation components (weave/trace and weave/flow).

---

All details are taken directly from the repository’s README; no additional features are inferred.