Marc Klingen is the co-founder and chief executive of Langfuse, an open-source AI engineering platform that helps developers trace, evaluate, and improve applications built with large language models. With fellow founders Max Deichmann and Clemens Rawert, he turned the difficulty of understanding unreliable AI agents into infrastructure for production observability, prompt management, and systematic evaluation.
Klingen studied business before pursuing computer science and innovation at the Technical University of Munich and the Center for Digital Technology and Management. His interests included open-source economics and blockchain oracle problems; earlier roles at Google, DHL, Cherry Ventures, and startups spanned software engineering, product management, and sales.
Building Langfuse
The founders developed Langfuse during Y Combinator’s Winter 2023 batch after discovering that impressive AI demonstrations revealed little about real-world behavior. Their initial product announcement described open-source product analytics for LLM applications that could connect production traces with quality, cost, latency, and user feedback.
Langfuse began on PostgreSQL, moved its core data layer to ClickHouse in late 2024, and introduced MIT-licensed product features and OpenTelemetry-based SDKs in 2025. ClickHouse acquired Langfuse in January 2026; Klingen announced the acquisition publicly, and the project retained its open-source, self-hostable foundations.
- Production traces reveal what agents actually do. Klingen uses runtime traces to identify unexpected customer needs, failed tool calls, incomplete retrieval instrumentation, and outdated agent instructions. He favors inspecting real executions before constructing elaborate evaluation systems.
- Agent skills must stay grounded in current documentation. Coding agents can recognize Langfuse from pretraining yet generate obsolete SDK integrations. Klingen’s approach to agent skills for reliable instrumentation gives them progressively disclosed guidance, current documentation, APIs, and CLI access. Agents should inspect CLI help instead of inventing flags, retrieve documentation as Markdown, establish the correct data region, and trace retrieval steps alongside model calls.
- Evaluation targets can erase important safeguards. Klingen applies LLM-as-a-Judge evaluations to repository changes, checking whether an agent implemented the intended observability. Optimizing solely for fewer agent turns, however, can eliminate documentation checks, human-approval gates, and links between prompt versions and production traces.
- Flexible infrastructure enables more autonomous improvement. Langfuse’s APIs let teams adapt tracing, evaluation, and prompt management to existing systems. Klingen is working toward agents that connect repositories, production traces, user feedback, and evaluation results while retaining human review where automated objectives miss essential context.
His AI Engineer Europe session on coding-agent skills also identifies unresolved practical problems: distributing skills, keeping installed guidance current, and deciding whether agents should optimize for rapid setup or deeper application-specific implementation.
