Langfuse vs W&B Weave (2026)

More than 100 integrations with agent frameworks, model providers and gateways send traces into Langfuse, data exports to PostHog, Mixpanel and blob storage, and prompt changes send notices through webhooks and Slack. These move telemetry in and data out, and there are no connectors that let an agent take authenticated actions in outside systems.

SourceLangfuse, langfuse.com homepage and pricingread 2026-09-21

Integrations with LLM providers, agent SDKs and harnesses such as the OpenAI Agents SDK, Claude Agent SDK and Google ADK, and frameworks such as LangChain, CrewAI and LlamaIndex let Weave trace calls, and it traces activity between MCP clients and servers, while alerts go out by Slack and email. These bring trace data in and send notifications out, and none connects an agent to a real system to take actions.

SourceWeights & Biases, docs.wandb.ai Weave index (integrations overview, agent integrations, MCP) and wandb.ai/site/pricingread 2026-09-21

For agents that run elsewhere, Langfuse provides tracing, evaluation and prompt management, and prompt composability links prompts together. There are no workflows that sequence, branch or retry an agent's steps, or that mix deterministic nodes with agent steps.

SourceLangfuse, langfuse.com homepage and pricingread 2026-09-21

Weave traces the customer's own multi-step and multi-agent workflows, including sub-agent delegations, and automations trigger actions from monitor metrics and trace activity. Weave itself runs no workflows that sequence, branch or retry steps, or that combine deterministic nodes with agent steps.

SourceWeights & Biases, docs.wandb.ai Weave index (trace sub-agents, automations)read 2026-09-21

Online evaluation runs LLM as a judge and code evaluators on production observations as they arrive, and monitors and alerts fire on metric and evaluator thresholds. Evaluation work therefore starts from incoming telemetry with no person starting each run, and prompt changes also emit webhooks.

SourceLangfuse, langfuse.com homepage and pricingread 2026-09-21

Monitors passively score production traffic, built-in signals score agent turns and tag quality and safety issues as they arrive, and automations trigger actions from monitor metrics and trace activity, with Slack and email alerts on the Pro plan. Scoring and actions start on incoming traffic without a person asking.

SourceWeights & Biases, docs.wandb.ai Weave index (monitors, signals, automations) and wandb.ai/site/pricingread 2026-09-21

Traces, prompts and evaluation datasets are stored, but there is no retrieval structure over the customer's documents or knowledge that grounds an agent's answers. Langfuse evaluates the grounding of RAG systems built elsewhere instead of providing one.

SourceLangfuse, langfuse.com pricing and llms.txtread 2026-09-21

RAG applications can be evaluated with LLM judges, and Weave traces the context a customer's retrieval returns. There is no document ingestion, index, retrieval layer or knowledge API that grounds an agent in company data.

SourceWeights & Biases, docs.wandb.ai Weave index (evaluate RAG applications)read 2026-09-21

Sessions, users and traces are recorded for inspection, and there is no session, conversation or long term memory that an agent reads back as context.

SourceLangfuse, langfuse.com pricing and homepageread 2026-09-21

Conversations, threads and turns of the customer's agent are recorded, and datasets, prompts and models are versioned as tracked objects. There is no session, workflow or long-term memory that an agent reads and writes.

SourceWeights & Biases, docs.wandb.ai Weave index (trace threads, track and version objects)read 2026-09-21

Protected deployment labels restrict who can promote a prompt version to production, and CI checks can require an approved baseline before a change ships, so changes to an agent's behavior pass a release gate that people control. There is no approval step, escalation rule or pause before an agent acts at run time.

SourceLangfuse, langfuse.com pricing and llms.txtread 2026-09-21

Guardrails run inline scorers on a user's input or a model's output in real time, before outputs reach users, and can block or modify a response when a score crosses a threshold, with built-in and custom scorers and AWS Bedrock Guardrails as examples. Every guardrail result is stored as a monitor. Annotation queues route traces to domain experts for structured feedback.

SourceWeights & Biases, docs.wandb.ai/weave/guides/evaluation/guardrails and the Weave index (annotation queues)read 2026-09-21

Reports for the SOC 2 Type II and ISO 27001 attestations are available from the Pro plan, and there is a region ready for HIPAA and a GDPR DPA. Access control covers RBAC at organization and project level, sign in with Google, Azure AD or GitHub, Enterprise SSO with Okta or Entra ID and SSO enforcement on the Teams add-on, SCIM provisioning, audit logs, data masking on the client side and data retention management.

SourceLangfuse, langfuse.com pricing and homepageread 2026-09-21

Users sign in with SSO through Google, GitHub or enterprise providers such as Okta and Azure Active Directory over OIDC, are organized into teams and projects with a restricted visibility scope, get role-based access at team or project level, automate with scoped service accounts, and are managed through the SCIM API. Multi-tenant Cloud and Dedicated Cloud are both SOC 2 Type II compliant with reports on the W&B Security Portal, Dedicated Cloud is HIPAA compliant, and paid plans include audit logs and customizable data retention.

SourceWeights & Biases, docs.wandb.ai/weave/guides/platform and wandb.ai/site/pricingread 2026-09-21

Hierarchical traces capture every LLM call, tool invocation and retrieval step of an agent run, so each run's steps and tool calls are visible. They appear as trace trees and agent graphs with session and user tracking, token cost and latency, and can be filtered by user, session, cost, latency or metadata. Audit logs are kept separately on Enterprise, data can be exported in batches or on a schedule to blob storage and to PostHog or Mixpanel, and historical data access runs 30 days on Hobby, 90 days on Core and 3 years on Pro.

SourceLangfuse, langfuse.com homepage and pricingread 2026-09-21

The Agents view shows conversations, turns, LLM calls, tool calls and sub-agent delegations with their cost, and the trace view follows nested execution paths. Calls can be filtered and exported through the Python SDK, REST API or UI, PII can be redacted from traces, and paid plans include audit logs and customizable data retention.

SourceWeights & Biases, docs.wandb.ai Weave index (agents view, trace tree, query and export calls, redact PII) and wandb.ai/site/weave and pricingread 2026-09-21

Customers of Langfuse Cloud choose US, EU or JP data regions, and there is a region ready for HIPAA and AWS PrivateLink. The MIT licensed platform self hosts with Docker Compose, Kubernetes via Helm, or Terraform on AWS, GCP and Azure, with paid Enterprise additions for self hosted deployments.

SourceLangfuse, langfuse.com pricing, homepage and llms.txtread 2026-09-21

Weave runs on W&B Multi-tenant Cloud in a North America region of W&B's Google Cloud account, or on W&B Dedicated Cloud on AWS, Google Cloud or Azure, with data in a dedicated ClickHouse cluster in the cloud and region of the customer's choice, IP allowlisting and private connectivity. It also runs on Self-Managed instances the customer hosts. W&B fully manages the two cloud options.

SourceWeights & Biases, docs.wandb.ai/weave/guides/platformread 2026-09-21

The Langfuse Assistant in the app is the one prebuilt agent, and a Langfuse skill packages best practice workflows for instrumentation, prompt management and API access into a customer's coding agent. Beyond those two, there is no set of prebuilt agents or templates for a customer to choose from.

SourceLangfuse, langfuse.com llms.txt and pricingread 2026-09-21

W&B Skills install into a coding agent to teach it to train models, build agents and analyze experiments on the W&B platform, a packaged, ready-to-install starting point for another agent. Weave has no ready-made agents, packaged employees or deployable workflows for a customer's own work. The Claude Code, Codex and OpenClaw plugins trace sessions and are integrations, not packaged agents.

SourceWeights & Biases, docs.wandb.ai/llms.txt (W&B Skills) and the Weave index, and wandb.ai/site/weaveread 2026-09-21

Customers configure their own LLM connections, which the Playground uses to test prompts on real production inputs and compare models side by side, and which power the LLM-as-a-judge evaluators the customer sets up. The customer brings its own providers and chooses the models Langfuse's own LLM features run on.

SourceLangfuse, langfuse.com homepage, pricing and llms.txt (LLM Connections)read 2026-09-21

Customers can connect OpenAI-compatible inference endpoints to a project as custom runtimes, managed from the UI or the Python and TypeScript SDKs, compare models in the Playground and Evaluation Playground, and run open source models through W&B Serverless Inference. Customers choose models and bring their own endpoints.

SourceWeights & Biases, docs.wandb.ai Weave index (custom runtimes, Playground, Serverless Inference)read 2026-09-21

An extensive public API has published rate limits by plan, alongside native Python and TypeScript SDKs, OpenTelemetry ingestion for Java, Go and other languages, a CLI, and a platform MCP server that lets IDE agents manage prompts and query traces. Stable APIs based on IDs manage evaluators, and there is an agent skill.

SourceLangfuse, langfuse.com homepage, pricing and llms.txtread 2026-09-21

Weave offers Python and TypeScript SDKs, a service REST API with an OpenAPI description, and an OpenTelemetry endpoint that accepts traces from any pipeline without the SDK. The W&B MCP server lets an IDE or agent query W&B data and documentation, W&B Skills teach coding agents to use the platform, and the Pro plan adds CI/CD automations.

SourceWeights & Biases, docs.wandb.ai/llms.txt and the Weave index (OpenTelemetry, query and export calls), and wandb.ai/site/pricingread 2026-09-21

Evaluation of the customer's application runs online and offline. It covers datasets, experiments from the SDK or UI compared side by side against baselines, LLM as a judge and code evaluators managed by API, custom scores, user feedback and human annotation queues, alerts on evaluator results, and CI checks that block regressions against thresholds and approved baselines. The customer's agent is tested with datasets before production, and output quality is scored over time.

SourceLangfuse, langfuse.com llms.txt, homepage and pricingread 2026-09-21

Evaluations run applications against versioned datasets with built-in, local and custom scorers, including LLM judges, for single-turn and multi-turn agents. Evaluations are compared side by side to spot regressions and ranked on leaderboards, monitors passively score production traffic, and the Pro plan adds CI/CD automations.

SourceWeights & Biases, docs.wandb.ai Weave index (evaluations, scorers, compare evaluations, monitors) and wandb.ai/site/weave and pricingread 2026-09-21

Langfuse observes and evaluates agents, and there is no browser, desktop or computer control by an agent.

SourceLangfuse, langfuse.com homepageread 2026-09-21

Weave covers tracing, evaluation, monitoring, guardrails, annotation and deployment, and no agent controls a browser, desktop or computer. The CoreWeave Sandboxes listed beside Weave are isolated environments for running agents, and they execute code instead of controlling a browser, desktop or computer.

SourceWeights & Biases, docs.wandb.ai Weave index and wandb.ai/site/weave and pricing; wandb.ai/site/pricingread 2026-09-21