Skip to content
Blog

LLM observability in practice

Practical writing on agent evaluation, release gating, and building AI features that stay reliable in production.

Regression testing LLMs vs APIs

Regression testing LLMs is different from regression testing APIs

API tests have a deterministic expected output. LLM output is probabilistic. Here is how to build a regression suite that accounts for output variance.

Jonas Park
Building eval suites that catch quality drops

Building eval suites that actually catch quality drops

A suite with ten metrics feels rigorous but can miss a real regression if thresholds are too loose. How to calibrate your suite against historical data.

Fenna van Dijk
Gating LLM releases like software

Gating LLM releases the same way you gate software releases

Software teams block a merge when tests fail. LLM teams should block a deploy when eval metrics regress. The mechanisms are the same, the tooling just needs wiring.

Sohrab Hosseini
Traces vs logs for agent observability

Traces vs logs: what you actually need for agent observability

Logging every LLM call to stdout gives you data. Tracing gives you the causal structure. The difference matters when an agent run fails after six steps and you need to find which one.

Jonas Park
Latency and quality tradeoffs in LLM evals

How to measure latency-quality tradeoffs in your eval suite

Switching to a faster model saves cost and latency. Does it also reduce output quality? How to use evals to answer that question before you ship.

Fenna van Dijk
Eval rubric for open-ended agent outputs

Writing an eval rubric for open-ended agent outputs

Deterministic metrics work for factual answers. When your agent writes prose or plans multi-step actions, you need a rubric. Here is a template that works in practice.

Sohrab Hosseini
Pilot to production scaling an LLM feature

Scaling an LLM feature from pilot to production

A pilot with ten users can hide failure modes that appear at a thousand. The observability gaps that kill LLM features at scale and how to close them before you launch.

Jonas Park
Structured outputs are not enough for reliable agents

Structured outputs are not enough for reliable agents

Constrained decoding guarantees your agent returns valid JSON. It does not guarantee the content of that JSON is correct or useful. What else you need.

Fenna van Dijk
A minimal eval baseline in one afternoon

A minimal eval baseline you can set up in one afternoon

You do not need a perfect eval suite to get value. Three metrics and a working threshold system is enough to catch regressions. A practical starting point.

Sohrab Hosseini
The cost of not knowing about agent regression

The real cost of not knowing your agent regressed

A three-week-old regression is not a fresh bug. The damage compounds every day before you find it. How to put a number on the cost of undetected LLM regressions.

Jonas Park
Why teams discover LLM bugs in production

Why most teams discover LLM bugs in production

The structural reasons engineering teams find out their AI feature broke when a customer tells them, and the tooling changes that break that pattern.

Fenna van Dijk

Put these ideas into practice

Orq gives you traces, eval scores, and CI gating in one place. Free to start, no card required.