Respan calls for proactive failure analysis
Respan argued teams must move beyond retrospective traces to proactive failure analysis, quality‑drift detection, and prescriptive next‑eval recommendations for complex agents—shifting observability toward prevention not just debugging. The pitch is for tooling that spots degradation patterns before customer impact. (x.com)
Respan unveiled its LLM engineering platform in a launch post and press coverage dated March 18, 2026, positioning the product as a “self‑driving observability” stack for production agents. (respan.ai) The platform stores every LLM interaction as a core data structure called a “span,” assembles spans into execution traces, and uses that trace data to trigger automated evaluations when prompts, workflows, routing logic, or model behavior change. (respan.ai) Respan’s company profile lists the platform as processing more than 1 billion logs and over 2 trillion tokens per month while claiming adoption by 100+ AI startups and enterprise teams and support for roughly 6.5 million end users. (ycombinator.com) The company announced a $5 million seed round led by Gradient with participation from Y Combinator and Hat‑Trick Capital on March 18, 2026. (tmcnet.com) A public Python package, respan‑tracing, exposes an SDK that automatically exports spans from instrumented apps, and Respan maintains a GitHub org with example projects for integrations. (pypi.org) A published customer story dated February 23, 2026 says Mem0 used Respan’s unified gateway and observability to improve memory retrieval accuracy and maintain 99.99% reliability at scale. (respan.ai) Respan describes a closed loop where traces feed evaluation scores (human and LLM judges), the platform surfaces regression signals and recommended “what to fix next,” and a single gateway can deploy the winning model/prompt configuration to production traffic. (respan.ai)
