Langfuse Lets You Watch AI Agents Work and Track Token Use Retroactively
The tool makes AI agent internals visible after the fact, turning a black-box token spend into a browsable record of every search, tool call, and model query.
Reporting from 1 source: GIGAZINE.
Langfuse is an open-source, self-hostable AI engineering platform that records the full execution path of LLM calls and AI agents as traces. From a browser, users can review which LLM was called, which tools executed, what was searched, processing time, token usage, and estimated costs. It integrates with OpenTelemetry, LangChain, the Google Gen AI SDK, and LiteLLM, and was part of Y Combinator's W23 batch.
The setup walkthrough starts with creating an organization and project, then issuing API keys. The writer notes a bug where the screen does not transition properly after project creation, advising users to click Projects in the side menu manually. The secret key appears only once, so it must be copied and saved at creation.
To demonstrate, the writer integrated Langfuse with Google Gemini through the Google Gen AI SDK and Observability for Google Gemini Models with Langfuse Integration. A deliberately redundant prompt asked the agent to re-check weather information with a different search keyword. The agent compared multiple sources and concluded a foldable umbrella was needed for Tokyo, despite clouds and sunny periods, because sudden rain was possible from mid-afternoon to evening.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.