The canon was written for a different machine
The single most useful thing you can notice about this literature is its dates: the guidelines everyone quotes were derived from recommenders, autocomplete and feed ranking, years before a model could call a tool.
The guidelines lesson gave you the eighteen and where they sit. This one gives you the observation that turns citation into a position, and it is not a clever reading. It is a list of products and a set of dates.
Table 2 of the CHI paper names the twenty products the eighteen guidelines were validated against, grouped into ten categories: e-commerce recommendations, navigation and route planning, music recommenders, activity trackers, autocomplete, social network feed filtering, email importance filtering, voice assistants, photo organisers, and web search. Read that list twice. Not one of those systems generates content, and not one of them takes an action in the world on the user’s behalf. Every one of them ranks, predicts, filters, or completes.
The paper was presented at CHI 2019, in Glasgow, on 4–9 May 2019. That is not an incidental fact about the eighteen. It is the fact that explains most of what they cover and almost all of what they do not.
The vocabulary gives it away before the dates do
You do not even need Table 2. The paper’s own framing does the work. Footnote 1, page 1, defines the object of study: “In this paper we use AI-infused systems to refer to systems that have features harnessing AI capabilities that are directly exposed to the end user.” And the introduction describes what those features do: automated inferences, “typically performed under uncertainty, often producing false positives and false negatives.”
False positives and false negatives is classifier vocabulary. It presumes a system whose output is a judgement that is either right or wrong, presented to a person who then does or does not act on it. That is a coherent and useful thing to design for, and the eighteen are good at it. It is also not what your review gate was solving, and it is nothing like a system that calls three tools, commits two side effects, and stops.
An agent that books a flight does not produce a false positive. It produces a booking. The failure is not a wrong label the user can ignore; it is a state change the user now has to unwind. No amount of careful reading gets that case out of a framework built on the other one.
The dates, so the claim is checkable rather than rhetorical
“This stuff is old” is a bad argument. It is dismissive, unfalsifiable, and it makes you sound like someone who has not read it. The precise version is a much better argument, and it takes four dates:
- 4–9 May 2019 — the eighteen guidelines are presented at CHI 2019.
- 6 October 2022 — ReAct (Yao et al.) is posted to arXiv, describing interleaved reasoning and acting: the loop pattern that essentially every tool-using agent now implements. The ICLR camera-ready version follows on 10 March 2023.
- 22 October 2024 — Anthropic ships computer use in public beta: a model “looking at a screen, moving a cursor, clicking buttons, and typing text.” Anthropic’s own description calls it “at times cumbersome and error-prone.”
- 25 November 2024 — the Model Context Protocol is announced: a standard for connecting assistants to the systems where data lives.
Three and a half years separate the guidelines from ReAct. Five and a half separate them from a model that can click things. Whatever the eighteen were derived from, it was not that.
Microsoft has not revised the eighteen. It has written around them
The obvious objection is that surely Microsoft updated the guidelines for generative AI and agents. It did not. The guidelines page still points at the 2019 paper and carries no revision date. What Microsoft did instead is publish separate documents, and the way they relate to the eighteen is genuinely interesting.
“Creating a dynamic UX: guidance for generative AI applications” on Microsoft Learn is the bridge document. Its page metadata gives an ms.date of 2024-09-16 and an updated_at of 2026-08-10. It links the HAX Toolkit explicitly and reuses eight guidelines verbatim under the same four lifecycle headings — G1 and G2 under first run, G5 and G6 during interaction, G9 and G11 when it is wrong, G15 and G17 over time. Then it adds three principles the eighteen do not contain: human in control, avoid anthropomorphizing copilot, and consider direct and indirect stakeholders. And one output-design tip whose heading is worth memorising because it inverts the reflex most designers were trained on: “Add appropriate friction (it’s a good thing!)”
Then compare it to “Human-centered design for agents”, the agent-specific page (ms.date 2026-05-20, updated_at 2026-07-14). Three things about it:
- Its lifecycle section runs First-run experience → During interaction → When the system is wrong — the HAX phase structure, almost word for word.
- It never mentions HAX or the eighteen guidelines. The structure was inherited without the citation.
- There is no “over time” section. The largest phase of the original framework — seven guidelines about memory, learning from behaviour, cautious updates, consequences and global controls — has no counterpart on the agent page, even though the 2024 gen-AI page did keep it.
One more detail before you use that page as evidence: its own metadata carries ai-usage: ai-generated. Microsoft is labelling it. Cite it as what it is — a published Microsoft position, generated with AI assistance — not as a considered research artifact the way the CHI paper is. There is also a human-authored companion, “When AI joins the team” by Dwyer, Basoah and Wang, published 27 January 2026 on Microsoft Design, carrying the same three principles.
Where people get burned
Watch how fast wording drifts once a guideline leaves its source. HAX G11 is “Make clear why the system did what it did.” Microsoft’s own gen-AI page renders it “Make it clear why the system did what it did.” Harmless in that instance, and a demonstration of the mechanism: each restatement is a little further from the text, and by the third hop somebody is confidently quoting a guideline that does not exist. Quote from the paper, not from a page quoting the paper.
Retrieval check
A designer says "the HAX guidelines are the standard for AI UX, so we should follow all eighteen for our agent." What is the accurate, non-dismissive response?
Check your answer
Something close to: they are the best-evidenced thing we have, and they were validated against twenty products in 2019 — recommenders, autocomplete, feed ranking, web search. Not one of them took an action. The authors say so themselves in section seven: the guidelines might not adequately address all types of AI-infused systems. So the ones about setting expectations and explaining behaviour transfer cleanly. The ones about dismissing and correcting a suggestion assume the suggestion has not already happened, which is the bit our agent breaks.
Three things make that work. It concedes the framework’s strength first. It cites the scope limit from the paper rather than asserting it. And it says which guidelines transfer, rather than treating the whole set as obsolete — because most of it does transfer, and claiming otherwise is the over-reach that gets you correctly argued down.
Hands on
Date-stamp every row in the position document
Done when: Every framework row in POSITIONS.md carries a provenance line with a publication date, a last-revision date or an explicit "no revision date published", and the class of system its evidence came from — each traceable to a page you opened yourself.
- Add a Provenance field to every framework row in
learning/agentic-ux-canon/POSITIONS.md: published, last revised, evidence base. - Fill the HAX row from the paper and the toolkit site. Published May 2019 at CHI; revision date — go and look, and write down what you actually find rather than what you expect to find. For the evidence base, write the ten product categories from Table 2 in your own shorthand.
- For every other framework row, open the publisher’s page and find the date. Where a page carries no date at all, write no date published in the cell. That is a finding, not a blank — an undated framework is a framework nobody has committed to maintaining, and you will want that fact later.
- Add one line under the HAX row: the single sentence you would say in a room to make the provenance point. Under thirty words, concedes the framework’s strength, and names the product categories rather than saying “it is old”.
- Bring the sentence into the chat. I will check whether it survives the obvious counter — “plenty of the eighteen still apply to agents, so what is your point?” — because that counter is correct, and a sentence that has no answer to it is not ready.
What this does not cover
This lesson establishes that the gap exists and where it came from. It does not say what falls into it. Naming the gap precisely is the whole of the next module: the suggest-versus-act lesson takes the guidelines that assume a suggestion can still be ignored and shows exactly which assumption an agent with tools breaks first, and the halfway-failure lesson takes the failure guidance — written for a wrong answer — and asks what your UI owes a user whose run stopped with two side effects already committed. Neither has a published answer, which is precisely why a position on them is worth having.
Read this next — primary source
Guidelines for Human-AI Interaction — Table 2 and Section 7 (CHI 2019, camera-ready)Amershi et al., Microsoft Research — free PDF, 13 pages; Table 2 is on page 5, the discussion on pages 11–12
You read this paper once already for the eighteen. Read it a second time for the two pages nobody quotes. Table 2 lists the twenty products the guidelines were validated against, by category — the whole provenance argument is in that one table, and seeing the list yourself is more convincing than being told what is in it. Section 7 is the authors stating their own scope limits: which systems the guidelines do not cover, and where they say specialised guidance would be needed. Having read the authors’ own caveat means your provenance argument is agreeing with them, not attacking them, which is a much stronger place to argue from.
Stuck, curious, or think this lesson is wrong? Ask your teaching agent. The lessons are the scaffold; the conversation is where the learning gets unstuck.