Your personal memory, across sessions, agents, and devices.

W&B Weave Explained: Agent Tracing, Evals and Agent Lens

memU Team memU Team

W&B Weave is the Weights & Biases toolkit for tracing and evaluating LLM applications and agents. CoreWeave bought Weights & Biases in May 2025. In September 2026 it launched CoreWeave Forge. The Weave docs now name a new product, CoreWeave Agent Lens, as Weave's successor for agent tracing. Below is how Weave traces and evals work and where Agent Lens fits, then why traces never reach the agent itself.

What is W&B Weave, in one sentence?

W&B Weave is an observability and evaluation platform that records each step of an LLM app or agent so developers can debug and score it.

That definition follows the Weave documentation, which now lives on CoreWeave's docs site. Python and TypeScript SDKs install with pip install weave or npm install weave. The open-source repository is Apache-2.0 licensed and shipped v0.54.0 on October 23, 2026.

How does Weave differ from experiment tracking and from Agent Lens?

Weights & Biases started with experiment tracking for model training. Weave covers the application side, after a model is deployed. Agent Lens is the newer production product.

W&B Models (experiment tracking) W&B Weave CoreWeave Agent Lens
What it watches Training runs and model checkpoints LLM calls, tool calls and functions in your app Production agent traces at large volume
Main unit A run with metrics A trace made of calls A conversation and its failures
Evaluation Metrics logged during training Datasets plus scorers, run on demand Failures turned into test cases for an LLM judge
Who it is for ML engineers Application developers Engineers plus product and domain reviewers
Status in October 2026 Part of CoreWeave Forge SDK still released, docs on CoreWeave's site Named the successor to Weave for agent tracing

Sources: the Forge launch release and the Agent Lens product page, plus the Weave docs.

How does Weave tracing work?

You decorate a function with @weave.op(). Each call to it becomes a record of its inputs and outputs. Nested calls form a trace, so a single agent turn shows as a tree from the model request down to each tool call. The docs say the Weave SDK is OpenTelemetry compatible and can trace agents down to single tool calls.

If you already emit OpenTelemetry spans, you can skip the SDK. The docs describe sending spans straight to Weave's OTLP endpoint. Teams with an existing tracing pipeline can add Weave as one more backend.

How do Weave evaluations work?

An evaluation in Weave has three parts. A dataset holds example inputs, often past failures, much like unit tests. Scorers are functions that grade an output, and a scorer can call a model to act as a judge. The model under test is a function or a model class with a predict method.

The evaluations guide shows the flow. You build an Evaluation object with the dataset and scorers, then call evaluate on your model. Each call is recorded as an evaluation run, with outputs and scores you can open row by row. Runs can repeat each example several times, which helps with nondeterministic models.

What changed with CoreWeave Forge and Agent Lens?

CoreWeave completed the Weights & Biases acquisition on May 5, 2025. On September 30, 2026, it launched Forge. CoreWeave calls it a development layer for building and improving models and agents. Forge combines W&B Models with OpenPipe's post-training tools and marimo notebooks. It all runs alongside CoreWeave's own services.

Agent Lens ships with Forge. Per its product page, it clusters failures across production traces and ranks them. Those failures become test cases scored by a tunable LLM judge. It offers PII redaction and bring-your-own-bucket storage. CoreWeave says Agent Lens detects 20 percent more critical failures at one-tenth the cost, measured against a general-purpose frontier LLM.

The Weave docs now call Agent Lens "the successor to Weights & Biases Weave". They also say both services already receive your agent tracing data, so there is nothing to migrate. The docs do not say Weave is deprecated, and the SDK still ships releases.

Should you start with Weave or Agent Lens?

For a new project that mainly needs development-time tracing and evals, start with Weave's SDK. Your traces reach Agent Lens anyway. If you run agents in production at volume and want failure clustering, try Agent Lens through Forge. Either way, instrument with OpenTelemetry so a rename or a product swap does not force you to re-instrument.

What do the vendor claims leave out?

The Agent Lens numbers come from CoreWeave's own benchmark against a general-purpose LLM. We found no independent test. The launch release does not mark which features are generally available, and it gives no per-edition pricing. The Weave docs use the word successor without a timeline. Before you build a long-term workflow on either product, check its support terms in your own Forge account.

The bottom line

Weave is still a working, documented way to trace and evaluate LLM apps. CoreWeave has named Agent Lens its successor but has given no timeline. Keep Weave evals as your regression suite, and revisit the choice when CoreWeave publishes a support plan.

Why don't agents learn from their own traces?

A trace in Weave is a record for a person. You open it to find the bad tool call, then fix the code. The agent that made the mistake never reads the trace. Its next session starts with the same context. It can repeat the mistake until someone ships a fix.

memU closes part of that loop for coding agents such as Claude Code and Codex. A scheduled background task slices each host's new session logs into jobs. The agent distills them into memory and skill Markdown, such as the test that needs a flag or the deploy step that failed. memU's MemoryService makes no LLM or chat calls. It stores and embeds what the agent wrote. A standing instruction in CLAUDE.md or AGENTS.md tells the agent to retrieve before answering. Traces remain the tool for finding bugs in the app you ship.

The hosted version is free and syncs across devices. Get an API key from memu.so, then send your agent one message:

Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is YOUR_MEMU_API_KEY.

The self-hosted install and the support matrix are in the memU GitHub repository.

FAQ

Is W&B Weave free?

The Weave SDK is open source under Apache-2.0. The hosted service now sits inside CoreWeave Forge, which has a free edition and paid tiers. CoreWeave has not published a per-edition feature list in its launch release.

Is W&B Weave being replaced?

The Weave docs call CoreWeave Agent Lens the successor to Weave for agent tracing. They do not announce a shutdown. Both services receive your tracing data.

Who owns Weights & Biases?

CoreWeave. It completed the acquisition on May 5, 2025, and W&B products are now part of CoreWeave Forge.

Does Weave work with OpenTelemetry?

Yes. The SDK is OpenTelemetry compatible, and you can send OTel spans directly to Weave's OTLP endpoint without installing the SDK.

What languages does Weave support?

Python and TypeScript. Both SDKs support tracing and evaluations, though some features, such as class-based scorers, are Python only.