Langfuse Co-founder Marc Klingen Demos Self-Improving Agent Stack

Marc Klingen, co-founder of the open-source observability platform Langfuse, shared a new paradigm for agent development at the 2026 AI Engineer World's Fair. He argued that while traditional agent optimization relied on humans reading traces weekly, modern approaches allow AI to handle most of the improvement loop.

The talk detailed a 'reference stack' that connects online tracing and monitoring with offline datasets, experiments, and evaluations. Klingen mapped layers of loops from next-token prediction up to human goal-setting, noting that AI is now capable of proposing fixes and maintaining datasets/eval criteria, while humans retain control over direction.

A specific demo in the video showed a coding agent using Langfuse skills to identify that its own changelog writer leaked internal jargon. The agent then updated the evals and back-tested a better version before shipping. This illustrated the use of implicit user signals like edits and approvals as feedback.

Klingen also emphasized the importance of owning your trace data layer, especially as agents read more of it. This resource is recommended for engineering professionals focused on LLM operations, evaluation framework design, and automated agent iteration workflows.