12 Comments
User's avatar
Raghav Mehta's avatar

The excerpt says you cannot read every output but still need to recognize the common ways the agent fails. How would you sample emails for that review so rare cases like long quoted threads or multilingual signatures actually show up instead of mostly clean ones? Stratifying the sample by thread length and language seems like a reasonable starting point.

Hamel Husain's avatar

Hi Ragav, Hamel author of the post here. Here’s a blog post that answers this question

https://hamel.dev/blog/posts/evals-faq/how-can-i-efficiently-sample-production-traces-for-review.html

Justin Lassiter's avatar

nobody wants to read the traces and it's the only thing that ever actually tells you what's wrong, like we had an agent summarizing client emails last week and the summary kept pulling in the signature block and nobody noticed for days because nobody looked. the three gulfs thing gives me a name for that at least, now I can say oh that's a specification problem and not just the AI being weird

People in Tech's avatar

The Three Gulfs framing is the part I will steal for my own work. Most eval talk jumps straight to scores, but the comprehension gulf explains why teams keep missing the same failures. Reading real traces first feels slow until it saves a whole rebuild.

Daniel Stein's avatar

The email example is brutal because 'summarize the key requests' sounds so specific until you ask how long the summary should be. I keep under-specifying stuff like that.

Tarun's avatar

I am student and literally yesterday we showed our mini project output to guide, we used llm and made it to refine text( formatting, grammar,

… ) but we didn't have any grounds to compare how well our project does the work, here this stack came at right time still a long way to go…

Nan Deng's avatar

The quoted-thread example makes the comprehension gulf pretty concrete: a prompt tuned on tidy standalone emails can end up summarizing the whole forwarded history instead of the new message. The specification gulf sits right next to it — "summarize the key requests" never says whether implicit asks count, or whether the sender is a display name or an address, so two runs can look confident and still disagree. I wonder whether the eval set keeps those as separate slices, because a long quote and a missing sender format are different failures.

Abhishek Pan's avatar

The fix for the sender-name bug is the part I'd underline: read it from the email header and stop asking the model. Before adding examples or fine-tuning for a generalization failure, it's worth checking whether that field needed an LLM at all. A lot of them don't.

People in Tech's avatar

The Three Gulfs framing is the part I will keep. Naming whether a failure is about understanding the data, stating the requirement, or generalizing to new inputs makes the next fix obvious instead of just trying another prompt.

Hamel Husain's avatar

Thanks for having me as a guest! I’ve also written this blog post that answers common questions I get about evals: https://hamel.dev/blog/posts/evals-faq/

Marcus Chang's avatar

The post describes the last step as verifying outputs. Once the agent calls tools that change data, I'd also check the records each run left behind. A support agent can write a perfectly reasonable reply to the customer and still have refunded the wrong order along the way.

For readers who want more on agent evaluation, I wrote a tutorial (1.2k+ GitHub stars): https://github.com/hardness1020/learn-agent-architecture/tree/main/sections/23-evaluation

ScuzzaMan's avatar

Whether you have a use case for AI Agents, and whether that Use Case has a Business Case, are the questions that should be answered before anyone thinks about how to put them into production. But nobody seems to be even asking these questions, let alone answering them.