AI Evals Explained · Aishwarya Srinivasan

How to Evaluate
AI Agents

A practical guide to AI evals: what they are, how they differ from benchmarks and classic machine learning evaluation, which metrics to track for which task, and the evaluation loop that separates teams shipping production agents from teams shipping demos.

Source: youtube.com/watch?v=_Er8Hao_gmQ · Channel: Aishwarya Srinivasan

Agenda

What We'll Cover

Understanding Evals

  • What is an AI eval?
  • Evals vs benchmarks vs classic ML evaluation
  • The metrics that matter — and the three families
  • A worked example: the demo generator

The Practice

  • Start with a golden dataset, not a metric list
  • Four ways to grade agent outputs
  • Evals are a loop, not a report card
  • Tools for running and monitoring evals
  • Key takeaways

Section 1 · What Is an AI Eval?

Is It Getting Better —
or Did You Just Break It?

An agent is a chain of decisions

Pick a tool → call it → read the result → decide what to do next — many places where things quietly go sideways. When it breaks in production you need to know where and why, not read transcripts one at a time at 2 a.m.

Lab scores run ahead of reality

Research shows roughly a one-third gap between how models score in the lab and how they perform in real, deployed work. Evals are how you close that gap before your users feel it.

"An AI eval is just a structured way of measuring how good your AI outputs really are." Every prompt change, model swap, or retrieval tweak gets the same question: did you help the agent or silently break something?

Section 2 · Evals vs Benchmarks vs Classic ML

Benchmarks Test the Model.
Evals Test Your System

Benchmarks

A generic test of the model.

The same test applied to every model: SWE-bench for coding, Tau-bench for customer-service agents, Terminal-Bench for command-line work, composite indexes that mash many together. Great for choosing which models to start with.

AI evals

A test of your specific agent.

Your task, your data, your prompts, your users — because the benchmark never saw your data, it can't tell you whether the model works for you.

A model can sit at the top of every leaderboard and still completely fall apart on your use case.

Use benchmarks to pick a starting model — then verify it with evals on your own task.

Section 2 · Evals vs Benchmarks vs Classic ML (cont.)

From a Multiple-Choice Test
to an Essay

Classic ML evaluation

Grading with an answer key.

Outputs were clean — a category, a number, yes or no. A test set with the right answers made grading objective: accuracy, precision, recall, F1. Everybody agreed on what "correct" meant.

AI evals

Grading without one.

Output is open-ended text or a whole sequence of actions — a hundred good summaries can look completely different from each other. There is no answer key: you need a rubric and somebody to use judgment.

That is the shift: we went from grading a multiple-choice test to grading an essay.

The good news: once the rubric mindset clicks, everything else gets simpler.

Section 3 · The Metrics That Matter

Four Metrics Cover Most Cases

Dozens of metrics feel like alphabet soup — in practice, only a handful are worth reaching for again and again.

Relevance

Did the output actually answer what was asked?

Faithfulness

Also called groundedness: is it true and backed by real data, or is the model making it up? This one is your hallucination check.

Correctness

Did it match the expected answer, when there is one?

Coherence

Does it actually read well?

These four are the top-level metrics — but the right metric for any given evaluation depends entirely on your task.

Section 3 · The Metrics That Matter (cont.)

Match the Metric to the Task

TaskWhat you care aboutGo-to metrics
SummarizationStayed true to the source without inventing · captured the main points · kept it tightFaithfulness · coverage · conciseness
ClassificationOutput is a clean label — old-school ML againAccuracy · precision · recall · F1
TranslationMeaning carried over · reads naturally in the other languageBLEU · COMET
RAGTogether they catch the two ways RAG breaks: bad retrieval or bad generationFaithfulness · answer relevance

Pick the metric after you know the task — never before.

Section 3 · A Quick Pause — Worked Example

One Slash Command → a Full Demo Package

Every weekly demo needs code, a README, a slide deck, a script, and a social post — so Srinivasan built a demo generator that produces all of it from a single slash command in her terminal.

Built on Mistral Vibe

  • Terminal-native coding agent powered by Mistral Medium 3.5
  • Understands a full codebase; helps write, test, refactor, deploy
  • Custom subagents: specialized agents with specific scope, tools, and permissions, delegated on demand
  • Custom skills: markdown files with YAML frontmatter declaring the slash command and workflow

The generate-demo skill

  • Asks clarifying questions first (built-in multiple-choice)
  • Research subagent gathers what is needed
  • Coding subagent writes the demo and runs it to confirm it works
  • Writer subagent drafts the README and slide outline
  • Writer reads a "voice file" of tone rules — output sounds like her, not generic AI copy

What used to take a full afternoon now takes a few minutes — composing specialized agents into one repeatable workflow.

Section 3 · Three Families of Metrics

Alphabet Soup, Sorted Into Three Families

1 · Overlap metrics

BLEU · ROUGE · METEOR

Measure word overlap against a reference answer. Fast, cheap, dead simple — but basically blind to meaning: the same idea in different words can be scored differently.

2 · Semantic metrics

BERTScore · BLEURT · COMET

Compare meaning using embeddings instead of matching words — they handle paraphrasing far better than overlap-based scores.

3 · Model-based metrics

LLM as judge · G-Eval

A model grades outputs against your rubric — the only family flexible enough to scale to open-ended, messy, reference-free agent outputs.

Agent outputs are open-ended and reference-free — so more and more, the thing doing the grading is a model.

Section 4 · Start With a Golden Dataset, Not a Metric List

Dataset First — Metrics Fall Out of It

What a golden dataset is

Not complicated: a handful of example inputs for your task, each paired with what a great output looks like and what a bad output could look like. Include your common cases, your annoying edge cases, and every failure you have already watched blow up — an answer key written for your own task.

Then work backwards

Ask of each example: what would make this output good, and what would make it bad? The quantified ways of judging those answers are your metrics.

Start small — about 52 examples is plenty Grow it over time
Do not start by grabbing a metric off a list. The dataset comes first, and the metrics fall out of it — not the other way around.

Section 5 · Four Ways to Grade Agent Outputs

Choose Your Grader

Human evals

The gold standard.

A domain expert scores outputs against your rubric. Slow, expensive, does not scale — save it for small, careful samples and treat it as the source of truth.

User feedback

Real ground truth.

Thumbs up or down, accept or reject, edits, task completion, repeat visits. Real people using your product — but noisy, and only available after you ship, so it cannot be your safety net.

Programmatic evals

Cheap and instant.

Code checks: valid JSON? the right tool with the right arguments? how long, how much? Run on every single change — but they only work when answers are clearly checkable, which pairs beautifully with classification.

LLM as a judge

Essay grading at scale.

A strong model grades outputs against your rubric — thousands of open-ended answers. The one rule: judge models have biases, so check their grades against a human grader on a sample first.

Judge tends to like longer answers Judge tends to like the first-seen response Validate the judge against a human on a sample — then it becomes your workhorse

Section 6 · Evals Are a Loop, Not a Report Card

Run It on Every Change, Forever

Not a one-time check before launch — the same idea as CI/CD: tests run on every change, forever. For AI agents: CI/CD plus continuous evaluation and continuous monitoring.

Decide what "good" even means Build the golden dataset Pick metrics backwards from the dataset Run a baseline across the whole dataset Group the failures: retrieval, formatting, edge cases Fix: prompts, retrieval, tool definitions Rerun the exact same dataset — nothing regressed? Ship and keep watching live traffic
You will never know whether your AI agent actually works by watching it succeed one time. You only know by measuring it — on your data, again and again.

Section 6 · The Loop in Production

Keep Going — Because Models Shift Under You

Providers can quietly swap models

Your model provider can update the models underneath you — and your agent's behavior can shift overnight without you touching a thing. That is why production monitoring never stops.

Failures flow back into the dataset

Real-world failures should go right back into your golden dataset — and then you run the whole loop again.

Teams that ship

Treat evals like continuous testing: always running, catching problems before users ever see them.

Teams that break

Run evals once at shipping time, then find out about every single bug from angry users.

Section 7 · Tools for Running and Monitoring Evals

Don't Build It All From Scratch

Running the evals

  • Promptfoo — test prompts and run evals right from the command line
  • RAGAS — RAG metrics straight out of the box

Tracing and monitoring

Watch what is happening in production with LangSmith, Langfuse, Arize, and Braintrust.

Promptfoo RAGAS LangSmith Langfuse Arize Braintrust

Pick the ones that fit your stack, wire them into the loop, and start measuring.

Key Takeaways

The Complete Checklist

Evals = structured, repeatable scoring: better, or broken?
Benchmarks test the model; evals test your system
Essays, not multiple choice: rubric + judgment
Core four: relevance, faithfulness, correctness, coherence
Task metrics: summarization, classification, translation, RAG
Three families: overlap, semantic, model-based
Golden dataset first — about 52 examples is plenty
Graders: humans, user feedback, code checks, LLM judge
Validate your LLM judge against a human sample
Evals are a loop: baseline, group failures, fix, rerun
Keep measuring in production — models shift overnight

The End

Keep Measuring Forever

Evals are what separate demo agents from shipped agents: decide what good means, build the golden dataset, let the metrics fall out, and rerun on every change — because "it works" is only true until you measure it.

Source: "AI Evals Explained | How to evaluate AI Agents?" — Aishwarya Srinivasan (youtube.com/watch?v=_Er8Hao_gmQ)

← → to navigate · swipe on mobile