W&B Weave

W&B Weave is the large language model and agent observability, evaluation, and guardrails product from Weights and Biases, the experiment tracking company that became an MLOps standard with more than a million users including OpenAI, Meta, and NVIDIA. Weights and Biases was founded in 2018 and was acquired by CoreWeave in a roughly $1.7 billion deal that closed in May 2025, with a stated interoperability pledge. Weave continues as the company's surface for teams building on top of models rather than training them from scratch.

Weave installs with a few lines of code, tracing any function through a simple decorator, from calls to OpenAI, Anthropic, and Google to retrieval, tool calls, and agent steps. It was built for multi turn, multi agent systems from the ground up, organizing traces into sessions, turns, steps, tools, and sub agents as first class concepts so a team can navigate a session the way the agent actually executed it. Built in and custom signals automatically capture and classify agent interactions to surface failure modes across millions of traces, and cost tracking records token usage and applies model pricing per call.

Evaluation is a core capability: a flexible framework with datasets, LLM as a judge scorers, and regression checks measures every change to prompts, harnesses, and models. Weave Guardrails add prebuilt scorers for safety and quality, including toxicity, bias, PII detection, hallucination, coherence, and relevance. Pricing in 2026 is a free tier, a Pro plan around sixty dollars a month, and custom Enterprise, sitting on usage meters for storage at three cents a gigabyte and Weave trace ingestion at ten cents a megabyte, with self hosted and VPC options and SOC 2, HIPAA, and ISO compliance. The main budgeting risk is trace ingestion overage on high volume applications, where an unprepared team can see a four figure surprise.