# Witdem > Witdem shows what AI workflows actually did and whether the run achieved the product goal. Witdem is runtime analytics for AI agents and multi step AI applications. It connects execution telemetry with application defined meaning. ## What Witdem does Witdem helps teams understand: * What happened during an AI run * Which models, tools, retrievers and execution paths were used * Whether a run completed * Whether the application achieved its product goal * How strong the evidence is for that conclusion * Where failures, retries and missing measurements occur * How time, token usage and cost relate to workflows, models and execution paths The central distinction is simple: > An AI run can finish successfully and still fail the job. A successful runtime does not automatically mean a successful product outcome. ## The Contract Witdem uses an application owned contract to define what success means. The contract can describe: * The result the application returned * The decision the system made * The product goal the run was expected to achieve * The checks and evidence required for assurance Witdem does not guess what business success means from a completed LLM call. The application defines the meaning. ## Runtime and product outcome Witdem separates two questions: ### What happened? Runtime telemetry describes the execution: * Model calls * Tool calls * Retrievers * Branches * Loops * Retries * Timing * Runtime completion ### What did it mean? Application semantics describe the outcome: * Product goal * Decision * Evaluation * Assurance * Application result This separation makes it possible to see when a system completed technically but did not deliver the intended result. ## Cost and usage Witdem connects measured spend with the run, workflow, model, provider and execution path that produced it. Cost and token data are only reported when sufficient measurement evidence exists. Missing data is not silently treated as zero. Provider reported cost takes precedence when available. This allows teams to investigate not only how much AI workflows cost, but why they cost that amount. ## Privacy Witdem is designed with a privacy first approach. Prompt and response capture can remain disabled. Structural runtime analytics can be collected without requiring the full content of conversations. Witdem can be self hosted, allowing organizations to keep their analytics infrastructure under their own control. Local analytics can use DuckDB. Witdem does not require teams to send application content to a third party merely to understand runtime behavior. ## Integrations Witdem is designed to work with existing AI stacks rather than replace them. Supported and documented integrations include: * Haystack * LangGraph * LangChain * OpenAI Agents * Anthropic * smolagents * LiteLLM * OpenTelemetry * Native Python Integration status can differ by framework and should be checked in the current documentation. ## OpenTelemetry OpenTelemetry can provide runtime telemetry such as execution and model information. Witdem can add application meaning on top of that telemetry through its contract and SDK integrations. OpenTelemetry can tell you what happened. Witdem can additionally show what that execution meant for the product goal. ## Self hosted Witdem can be deployed as a self hosted analytics stack. The system is designed to fit into an existing application architecture and observe real execution without requiring the application to be rebuilt around Witdem. ## Who Witdem is for Witdem is built for teams operating AI agents and AI applications in production. Typical users include: * AI and ML engineers * Application engineers * Platform and infrastructure teams * Product teams responsible for AI applications * Enterprise teams responsible for reliability, cost and governance ## Important distinctions Witdem is not an AI orchestrator. Witdem does not replace the framework running the application. Witdem does not decide what a successful business outcome means. Witdem does not silently convert missing measurements into zero. Witdem does not rely on a hidden model to invent a business success definition. ## Core terminology **Run** One correlated execution of an application. **Trace** Telemetry describing what happened during execution. **Contract** The application's explicit definition of what the result should mean. **Product goal** The outcome the application is supposed to achieve. **Assurance** The evidence supporting the conclusion that the product goal was achieved. **Runtime completion** Whether execution completed. **Not measured** A required measurement is unavailable or insufficiently evidenced. ## Key idea > Witdem shows not only whether the system ran, but whether the run achieved the outcome the application defined. ## Resources - Website: https://witdem.com/en - Documentation: https://docs.witdem.com/ - Getting started: https://docs.witdem.com/getting-started/ - YAML contracts: https://docs.witdem.com/contract-tutorial/ - Concepts: https://docs.witdem.com/concepts/ - Open-source repository: https://github.com/ebrahimisoheil/witdem-oss - Python SDK (PyPI): https://pypi.org/project/witdem-sdk/ - Analytics package (PyPI): https://pypi.org/project/witdem-analytics/ - Guide (EN): https://witdem.com/en/agentic-workflow-observability - Guide (DE): https://witdem.com/de/agentic-workflow-observability - Haystack observability (EN): https://witdem.com/en/haystack-observability - Haystack observability (DE): https://witdem.com/de/haystack-observability - LangGraph observability (EN): https://witdem.com/en/langgraph-observability - LangGraph observability (DE): https://witdem.com/de/langgraph-observability - LangChain observability (EN): https://witdem.com/en/langchain-observability - LangChain observability (DE): https://witdem.com/de/langchain-observability - OpenAI Agents observability (EN): https://witdem.com/en/openai-agents-observability - OpenAI Agents observability (DE): https://witdem.com/de/openai-agents-observability - Document intake case study (EN): https://witdem.com/en/document-intake-case-study - Document intake case study (DE): https://witdem.com/de/document-intake-case-study - Witdem and Langfuse (EN): https://witdem.com/en/witdem-and-langfuse - Witdem and Langfuse (DE): https://witdem.com/de/witdem-and-langfuse - Witdem and LangSmith (EN): https://witdem.com/en/witdem-and-langsmith - Witdem and LangSmith (DE): https://witdem.com/de/witdem-and-langsmith - Witdem and Arize Phoenix (EN): https://witdem.com/en/witdem-and-arize-phoenix - Witdem and Arize Phoenix (DE): https://witdem.com/de/witdem-and-arize-phoenix - Tracing, evaluation, and product analytics (EN): https://witdem.com/en/tracing-evaluation-product-analytics - Tracing, evaluation, and product analytics (DE): https://witdem.com/de/tracing-evaluation-product-analytics