Max Deichmann, co-founder of Langfuse (acquired by ClickHouse), shares how his team built the open-source LLM engineering platform that became the observability layer for production AI agents and why the engineer's future job looks less like coding and more like managing a decision inbox.
The conversation covers the full stack of building and operating AI agents in production: pre-production testing, tracing, online evaluation, and the iteration loop that separates teams who are actually succeeding with agents from those chasing the hype. Max is unusually honest about where Langfuse itself still falls short, and what agentic workflows look like in practice, including copying 3am Slack alerts into Codex and waiting for the agent to start the investigation.
From the YC pivot that started it all to the half-page doc that led to the ClickHouse acquisition, this episode gives CTOs a grounded, no LinkedIn-filter view of where AI agents actually stand today.
Key topics:
- Why traditional observability tools fail for non-deterministic AI applications
- The Langfuse loop: pre-production testing, tracing, online evaluation, and iteration
- Open source as a go-to-market strategy: adoption without a sales team
- The ClickHouse acquisition: the two reasons that fit on half a page
- Agentic on-call: how Max's team handles incidents with Codex today
- The real state of agents in production, and where the hype still outpaces reality