Evals
Long-horizon agent benchmarks are fragmenting: a field guide to what each one actually measures
Originally published on the Arize AI blog: Long-horizon agent benchmarks are fragmenting: a field …
AI evals are just testing (with a much weirder answer key)
Years ago, before “AI” meant chatbots and before anyone said the word “eval” …
The Phoenix Project still holds up, even if you replaced all the code with agents
I reread The Phoenix Project last month. I do this every year or two — it’s one of those books …
Train, dev, test: the split that makes an LLM judge trustworthy
Say you’ve built an LLM judge and you want to know if it’s any good. The obvious move is …
Evals do three jobs, not one
Ask most people what an eval is for and you’ll get some version of “testing”. You …
Anatomy of an evaluator: what happens when your prompt meets your traces
Last time I pulled apart the eval prompt - the role, criteria, rubric and examples you write to tell …
Anatomy of an eval prompt: what to actually put in it
When people decide to use an LLM as a judge, the prompt they reach for first is almost always some …
Spans, traces and sessions: the three zoom levels of an AI app
The moment you start tracing an AI app, three words turn up everywhere: span, trace, session. People …
Evals aren't a step at the end. They run the whole way through
There’s a version of building an AI app that goes like this. You build the thing, you get it …
Four ways to run an eval, from a cheap unit test to a full-blown agent
Someone asked me last week how you actually run an eval on an AI app. I gave the honest answer, …
How to build an eval you can actually trust
Here’s how most people build an eval. They open a file, write an LLM judge prompt that says …