Both trace agents, run offline experiments and score production traces with LLM judges. Fiddler is built for production monitoring of ML models and agents, with guardrails that can block inline through an AI gateway; Weave is built for the development loop, with prompt versioning, a playground and review queues next to W&B model training.[1][6]
Fiddler and W&B Weave both reach beyond LLMs, but in opposite directions: Fiddler monitors classic ML models and LLM agents in production and adds real-time content guardrails, while Weave sits next to Weights & Biases model training and adds prompt and evaluation tooling. CoreWeave acquired W&B in May 2025 and has sold Weave inside CoreWeave Forge since September 2026. Weave has a prompt playground and annotation queues; Fiddler has neither.
| Decision factor | Fiddler | W&B Weave |
|---|---|---|
| Ownership | Independent; $30M Series C in January 2026 | CoreWeave since May 2025; sold inside CoreWeave Forge since 30 September 2026 |
| Beyond LLMs | Monitoring for classic ML models in production | W&B model training and experiment tracking |
| Runtime guardrails | Partial: real-time content scoring; inline blocking through a third-party AI gateway | Partial: scorers called from Python code can block or modify inputs and outputs |
| Prompt tooling | None documented | Versioned prompts and a playground |
| Human review | Span annotations in public preview; no review queue | Annotation queues with reviewer fields |
| Self-hosting | VPC or on-premises on Enterprise | Kubernetes with a Weave licence and a W&B Platform install |
| Entry pricing | Developer: $0.002 per trace, SaaS | Forge Free: 1 GB of Weave ingestion a month; Pro from $60 a month |
| Capability | Fiddler | W&B Weave | Prefactor |
|---|---|---|---|
| Observe | |||
| LLM call tracing | ✓ | ✓ | ✓ |
| Multi-step agent tracing (tool calls, nested spans) | ✓ | ✓ | ✓ |
| Cost tracking | Partial | ✓ | ✓ |
| OpenTelemetry support | ✓ | ✓ | ✓ |
| Evaluate | |||
| Offline evals (datasets, experiments) | ✓ | ✓ | ✓ |
| Online evals on production traffic | ✓ | ✓ | ✓ |
| LLM-as-a-judge scoring | ✓ | ✓ | ✓ |
| Code-based evaluators | ✓ | ✓ | ✓ |
| Human review and annotation | Partial | ✓ | ✓ |
| Prompt versioning and playground | — | ✓ | — |
| Act on findings | |||
| Runtime controls on live agent actions | Partial | Partial | ✓ |
| Gateway or proxy in the request path | — | Not stated | — |
| Licence and deployment | |||
| Open source | Partial | Partial | Partial |
| Self-hosting | ✓ | ✓ | Partial |
Fiddler started in ML model monitoring and still covers it: one platform watches classic models, LLM applications and agents in production. It is an independent company that raised a $30M Series C in January 2026, bringing its total funding to $100M, and it describes itself as a control plane for AI agents.
Weave starts from the other end of the model lifecycle. It is the LLM and agent toolkit of Weights & Biases, the training and experiment-tracking platform that CoreWeave acquired in May 2025. Since 30 September 2026 it has been sold inside CoreWeave Forge alongside W&B Models, a registry and CoreWeave inference, and W&B hosted accounts sign in through Forge. CoreWeave now recommends Agent Lens, a separate product in public preview, for agent tracing.
So the first question is which team owns this purchase. A production ML monitoring team will find Fiddler familiar. A research or applied ML team that trains in W&B will find Weave familiar.
Both offer content checks during a run, wired in differently. Fiddler Guardrails score requests and responses in real time for faithfulness, personal and health data, secrets and harmful content, using Fiddler's own Centor models, and return scores for your application to act on. Plugged into a third-party AI gateway such as Kong, AgentGateway, LiteLLM or Portkey, they can block or redact inline without code changes.
Weave's guardrails are scorers you call from Python application code; they can block or modify inputs and outputs before they reach users. They need code changes, and the TypeScript SDK does not support them. Neither product documents a human approval step or policies on an agent's tool calls.
This is where the development-loop difference shows. Weave has versioned prompt objects, a playground for editing prompts and comparing models side by side, and annotation queues that route traces to reviewers with structured fields and export the results to datasets. Its Evaluation API compares runs and publishes leaderboards.
Fiddler documents no prompt versioning or playground; its prompt specs are templates for LLM judges, not application prompts. Human review is span-level annotation, in public preview since August 2026, without a queue or labelling workflow. Its evaluation strengths are elsewhere: 13 pre-built evaluators, evaluator rules that score new production spans with backfill and downsampling, and Fiddler-hosted judge models, including a Ministral 3 8B judge made generally available in September 2026. Teams that want to keep judge spend down can run evaluators on Fiddler's models rather than their own provider keys.
Cost tracking differs too. Weave calculates token costs automatically for supported providers, including cache tokens, but only in Python. Fiddler tracks token usage and stores monetary cost when the trace source emits cost attributes, for example from LiteLLM.
SDK coverage leans Python on both sides. Fiddler's native agent SDKs are Python only, for LangGraph, LangChain, Strands and Google ADK, with OpenTelemetry for other languages. Weave has Python and TypeScript SDKs, with fewer features in TypeScript.
As of 20 October 2026, Fiddler's self-serve Developer plan is $0.002 per trace on SaaS, adding observability, experiments, custom evaluators and single sign-on. A free plan covers guardrails only and starts with a demo request, and VPC or on-premises deployment is Enterprise, on a quote. Weave is priced through CoreWeave Forge: Forge Free includes 1 GB of Weave ingestion a month and up to 5 model seats, Forge Pro starts at $60 a month with 1.5 GB for teams under 50 employees, and extra ingestion is listed at $0.10 per MB.
Fiddler's unit is the trace, whatever its size; Weave's is the byte, whatever the trace count. Teams with long, tool-heavy traces may prefer the first; teams with many short calls, the second. Compare Fiddler's pricing and CoreWeave Forge pricing on your own numbers.
We sell the layer this section describes. Read it with that in mind.
Stated plainly, since Prefactor publishes this page: Prefactor is an agent evaluation platform that scores each step of a run and acts on what it finds through runtime policies on tool calls and data access, approval holds for risky actions and a kill switch for a run, an agent or a team. Fiddler and Weave both check content during a run, through a gateway or Python code, but neither holds an agent's action for a person to approve. Prefactor does not monitor classic ML models, has no prompt playground or prompt registry, and its platform is hosted, so teams that need those should choose Fiddler or Weave. See the full comparison of 15 platforms.
Reviewed against public product and documentation pages on October 20, 2026. If a vendor has changed a feature, product name, or positioning since then, send a correction and we will update it. Numbered source links in the page body point to the ordered sources below.
Prefactor scores each step of every agent run and can hold or stop an action when a score says it should not happen. Start on the free tier: 5,000 spans a month, no credit card required.
Prefactor helps teams observe, evaluate, and improve their AI agents in production, across every framework and provider.