FROM THE AI FRONTIER
Frontline news — November 19: choose the next step before asking for text
One AI Engineer talk: how choosing among options differs from generating an answer, with an exercise using fictional enquiries.

If a task comes down to choosing from existing options, define those options and the checking rule first. Not every step needs a new extended answer. In today's edition, I examine one AI Engineer recording from the past two weeks and suggest an exercise in sorting enquiries. We can save the elegant essay for tasks that actually need an essay.
Video: choose first, then generate Original: Your Agent Doesn't Need an LLM for Every Decision. Channel: AI Engineer. Speaker: Raymond Weitekamp, OpenProse. Language: English. Duration: 51:48. Recording published: November 11, 2026, at 02:00:07 Moscow time. YouTube metadata gives November 10, 15:00:07, UTC−8. This is the recording's publication date, not a verified date of the talk or a product launch.
At 2:07–2:14, the speaker distinguishes the model's judgment from the program's decision. The model supplies a judgment; code determines what to do with it. At 9:21–9:31, he describes an approach where calling a language model or an agent is only one possible next step. This is the speaker's position and an architectural description, not proof of superiority over a conventional chatbot.
:::note What was actually examined
The official transcript passages 0:00–2:42 and 6:25–9:48, plus the related introduction, were read. The entire 51-minute talk was not reviewed. Automatic captions may contain errors; the speaker's name was checked against the official page.
:::
Practical value: separate choosing from writing Imagine a learning exercise: you need to sort visitors' enquiries. One person wants booking information, another asks about a refund, and a third has a complaint. I suggest assigning a category first and drafting a reply separately. This is an editorial example, not a service demonstrated by the speaker or a result from our testing.
- Define categories: “booking”, “refund”, “complaint” and “unclear”.
- Describe the boundaries. For example, dissatisfaction with a delayed refund may fit two categories. Allow “unclear” for this situation in advance.
- Ask for one category and a short passage from the original message as evidence.
- Check the choice against the message. Only then move on to a draft reply.
You do not need to build a dedicated agent to try this separation in a chat. However, a chat exercise does not reproduce the talk's architecture or verify its speed, cost or accuracy.
:::summary An exercise without automatic sending
Use fictional enquiries and label them yourself. Then give the same set to the assistant, compare categories and examine disagreements. Do not judge the result only by how convincing its explanation sounds.
:::
Where the conclusions stop The passage examined does not establish that a separate decision model is always cheaper, faster or more reliable than text generation. Those properties need separate measurement for your task. This edition does not recommend installing a particular product, provide pricing or promise that the approach is available in every assistant.
My editorial takeaway: before asking “write a reply”, consider whether you need text now or first need to choose the next step. Start with clear options and permission to acknowledge uncertainty.
Sources
- AI Engineer — original recording
- AI Engineer — official timestamped transcript
- AI Engineer — official talk page