<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Cinder — engineering blog</title><description>Cinder is the observability layer for AI-native backends. Trace model calls, tool invocations and the services around them in one timeline.</description><link>https://cathode.sandbox.koalastudio.dev/</link><language>en-us</language><item><title>What a trace should tell you about an agent run</title><link>https://cathode.sandbox.koalastudio.dev/blog/what-a-trace-should-tell-you/</link><guid isPermaLink="true">https://cathode.sandbox.koalastudio.dev/blog/what-a-trace-should-tell-you/</guid><description>Most agent traces record what happened. The useful ones record what was decided, and on what evidence. Here is the difference, in spans.</description><pubDate>Tue, 14 Jul 2026 00:00:00 GMT</pubDate></item><item><title>Sampling without losing the incident</title><link>https://cathode.sandbox.koalastudio.dev/blog/sampling-without-losing-the-incident/</link><guid isPermaLink="true">https://cathode.sandbox.koalastudio.dev/blog/sampling-without-losing-the-incident/</guid><description>Head-based sampling drops the traces you needed. Tail-based sampling costs what you were trying to save. We ended up somewhere in between, and the details matter.</description><pubDate>Tue, 23 Jun 2026 00:00:00 GMT</pubDate></item><item><title>Model-graded evals are still evals</title><link>https://cathode.sandbox.koalastudio.dev/blog/model-graded-evals-are-still-evals/</link><guid isPermaLink="true">https://cathode.sandbox.koalastudio.dev/blog/model-graded-evals-are-still-evals/</guid><description>If a model is grading your output, the grader is a dependency with a version, a failure mode and a drift profile. Treat it like one.</description><pubDate>Thu, 28 May 2026 00:00:00 GMT</pubDate></item></channel></rss>