Aakash GuptaHow to Build AI Evals Step-by-Step | Daniel McKinnon | Product Growth
At a glance
WHAT IT’S REALLY ABOUT
Building agentic AI evals with domain expertise and scoring rigor
- Evals are a practical way to specify and communicate expected AI product behavior through concrete examples, often covering much of what traditional PRDs tried to capture in prose.
- Offline evals use a fixed, representative prompt set with scoring as a proxy for user satisfaction, helping teams learn feasibility and avoid shipping broken experiences to production.
- Classic QA benchmarks (e.g., MMLU) are increasingly saturated, so modern evals must focus on agentic, multi-step tasks with tool use and longer horizons.
- A workable eval set needs “Goldilocks” difficulty (not 0% or 100%) so teams can improve models/harnesses over time and know when an eval has hit its ceiling.
- Deep subject-matter expertise is the bottleneck: building reliable agentic evals depends more on understanding the domain and success criteria than on having a fancy template or tooling.
IDEAS WORTH REMEMBERING
5 ideasTreat evals as the core spec for AI behavior—examples beat prose.
McKinnon argues most PRDs over-index on describing behaviors in text, which breaks down for generative systems; curated input–output examples (and scoring) communicate “what good looks like” more precisely to engineering and research.
Start offline to validate feasibility before exposing users to failures.
Offline evals act as pre-production proxies using a pre-baked dataset of prompts and an answer key/scorer; they help you learn early whether you need a different model, a better harness, or a different product approach.
Design your eval set to land around 25–50% success initially.
If an eval is too easy (near 100%), teams can’t measure progress; if it’s too hard (near 0%), you can’t tell if it’s feasible. A mid-range baseline provides room for optimization and clearer signals.
Modern evals must measure multi-step, tool-using agent behavior—not just QA.
As labs shift from search/QA narratives to agents (coding, computer use, knowledge work), evals must score end-to-end task completion across long horizons, where partial correctness and tool choices matter.
Scoring is the hardest part of agentic evals—optimize for automation.
Human review can work for many QA tasks, but agentic workflows produce long traces and complex outcomes; scalable iteration requires automatic scoring (often via programmatic checks, structured outputs, or LLM judges).
WORDS WORTH SAVING
5 quotesThe best way to actually communicate what the product should do is through examples, and that's what an eval is.
— Daniel McKinnon
You must deeply understand the problem.
— Daniel McKinnon
With agentic work, it's much more challenging because the time horizon tends to be very long, and the final output is a collection of many, many, many steps that it took.
— Daniel McKinnon
When you're creating evals like this, it is literally just prompts, responses, and ways of scoring whether response is correct.
— Daniel McKinnon
The Metas and the Googles and all of the other large companies have to reinvent themselves right now in the age of AI. Every single PM is gonna start building AI features. You cannot exist as a PM without understanding this.
— Daniel McKinnon
High quality AI-generated summary created from speaker-labeled transcript.
