Devcon 8 · Agentic Hub · 4 Nov 2026

Your next customer won’t click.

Designing for humans and agents

Fortunato Cinquepalmi Product Lead, Kleros

→ GET /talk Accept: text/markdown

← 200 text/markdown

[0:30]

On screen: the title. A cursor drifts toward "click." and slows down. Let it drift while you start. "Hi, I'm Fortunato. I lead product at Kleros. This talk is about a customer that doesn't use your buttons."

Press: a quiet request goes out at the bottom right, and the answer comes back: 200, Markdown, a few kilobytes. The cursor is still not there. "While the cursor is still on its way, that request already went out and came back. No click. That is a customer too. So today: how do you design for people and for agents at the same time?"

Careful: don't say people stop clicking, or that the website is dead. The claim is smaller: some of your next customers arrive as agents, by text. The bytes on the line are the real size of this talk's Markdown door (measured live on the laptop, or at build time).

The whole talk in one picture

The agent’s channel is getting wider. The human’s is getting narrower.

What your agent takes in

Every departureEvery seat classThe price rulesThe timetable The docsThe skill fileThe termsThe refund rules The ID rulesStep-free accessThe other optionsYour budget The APISeats leftCancellationYour mission

What reaches you

Your agent asks

Pay 12.80 USDC?

4 ferry tickets · Sat 7 Nov

Not nowYes, pay

Design what crosses the narrow one.

[0:45]

On entry: the two lines, then the left side fills with what the agent takes in, and one card appears on the right. "Here is the whole talk in one picture. When an agent buys something for you, it takes in everything: every departure, every price rule, the docs, the terms, the refund rules. That side keeps getting wider. And what reaches you, the person? One card. Four ferry tickets, 12.80, yes or not now. That side keeps getting narrower, because the agent does more of the work."

Press: the left side steps back, the card lights up, and the last line rises. "So the design job changes. Design what crosses the narrow one: what you asked for, the one yes, and the receipt after. Get those few things right, and give the agent its own door for everything else. That is this talk."

Careful: this is my model, not a measurement. The only number is the demo's own price. Don't say people are leaving the loop: the narrow channel never closes.

One door, built for eyes.

What a person sees

The Harbour Line home page: menu, sign in, a harbour picture with a ferry, a Book tickets button, and a cookie banner

What an agent gets

the same page, as HTML28,237 bytes

<div id="app" inert=""><header class="top"><div class="top-in"><button type="button" class="brand" data-act="nav.home" aria-label="Harbour Line home"><svg class="mark" width="32" height="32" viewBox="0 0 32 32" aria-hidden="true"><circle cx="16" cy="16" r="15.5" fill="#0A3C40"></circle><circle cx="16" cy="16" r="10.5" stroke="#FBF6EC" stroke-width="6" fill="none"></circle><path d="M25.74 19.93 A10.5 10.5 0 0 1 19.93 25.74" stroke="#B5462F" stroke-width="6" fill="none"></path><path d="M12.07 25.74 A10.5 10.5 0 0 1 6.26 19.93" stroke="#B5462F" stroke-width="6" fill="none"></path><path d="M6.26 12.07 A10.5 10.5 0 0 1 12.07 6.26" stroke="#B5462F" stroke-width="6" fill="none"></path><path d="M19.93 6.26 A10.5 10.5 0 0 1 25.74 12.07" stroke="#B5462F" stroke-width="6" fill="none"></path><circle cx="16" cy="16" r="13.6" stroke="#FBF6EC" stroke-opacity=".35" stroke-width=".8" fill="none"></circle></svg><span class="wordmark">Harbour<span class="wm-line">Line</span></span></button><nav class="top-nav" aria-label="Main"><span>Timetable</span><span>Fares</span><span>Island guide</span><span>Help</span></nav><div class="top-user"><button type="button" class="link-btn" data-act="nav.signin">Sign in</button></div></div></header><main id="view" class="view-home"><section class="hero"><svg class="hero-svg hero-wide" viewBox="0 0 1200 460" preserveAspectRatio="xMidYMax slice" aria-hidden="true"><rect x="0" y="0" width="1200" height="301" fill="url(#hl-sky)"></rect><circle cx="930" cy="214" r="120" fill="url(#hl-halo)"></circle><circle cx="930" cy="214" r="44" fill="#F7E3BA"></circle><g class="gulls" fill="none" stroke="#F4E6CC" stroke-width="2" stroke-linecap="round"><path d="M650 120 q8 -8 16 0 q8 -8 16 0"></path><path d="M700 146 q6 -6 12 0 q6 -6 12 0"></path><path d="M1070 132 q6 -6 12 0 q6 -6 12 0"></path></g><path d="M640 300 C700 268 760 238 830 236 C880 234 905 258 950 262 C1000 266 1030 240 1080 242 C1130 244 1170 270 1200 282 L1200 300 Z" fill="#4F8C85" opacity=".75"></path><path d="M720 300 C770 284 806 264 858 263 C904 262 934 282 984 284 C1042 286 1090 268 1140 270 L1200 276 L1200 300 Z" fill="#2C6F6A"></path><g transform="translate(1104 280)"><path d="M0 14 L22 14 L18 19 L4 19 Z" fill="#1D5C59"></path><path d="M11 -4 L11 13 L22 13 Z" fill="#F4E6CC" opacity=".9"></path></g><rect x="0" y="300" width="1200" height="160" fill="url(#hl-sea)"></rect><rect x="0" y="299" width="1200" height="2" fill="#F3DDB5" opacity=".55"></rect><rect x="870" y="306" width="120" height="3" rx="1.5" fill="#F7E3BA" opacity=".55"></rect>

28,237 bytes of HTML1,004 bytes are words a person reads

Harbour Line: the made-up ferry company in our demo. Page as Chrome draws it, measured 4 Nov 2026.

[1:00]

On screen: a ferry company's home page. It's Harbour Line, the made-up ferry company you'll see again in the race. "This is the web we built for the last twenty years. You read the page."

Press: small tags point at the parts. "A menu. Sign in. A big picture. A button. And before anything else, a cookie banner. Every piece is made for eyes and a hand on a mouse."

Press: the right side fills with code. "Now the same page, the way an agent gets it. This is the HTML of that home page: about 28 thousand bytes of tags, classes and drawing instructions."

Press: the words light up, and the bar at the bottom shows how much of it they are. "The words a person actually reads are about a thousand bytes of it. The rest is there to draw the picture. One door, built for eyes."

Careful: this is our own demo page, measured on this laptop this morning. It is not a study of the web, so don't generalise the ratio. Don't call it a real company.

Same link, two doors.

GET/talk

Live from this laptop

For a person Accept: text/html

—

For an agent Accept: text/markdown

—

Cloudflare measured its own launch post: 16,180 tokens as a web page, 3,150 as Markdown, about 80% less. blog.cloudflare.com/markdown-for-agents · Feb 2026

[1:15]

On screen: one link, /talk, served by this laptop. The tag at the top right says if it is live or recorded. "So what if the same link had two doors? This is the page for this talk, on the laptop in front of me."

Press: the left door opens. The slide asks for /talk the way a browser does, Accept text/html, and shows the answer: status, type, size, and the page itself. "Ask it like a browser and you get the web page."

Press: the right door. Same link, but the request says Accept text/markdown. The terminal prints the real reply: 200, text/markdown, the size, and a Link header pointing to llms.txt. Then the Markdown itself. "Ask it like an agent and you get the talk as text. Same address, no scraping, nothing to guess. And the Link header tells it where the instructions are."

Press: the two sizes as bars, and Cloudflare's number underneath. "Here, the Markdown is a small fraction of the page. Cloudflare measured the same thing on its own launch post: 16,180 tokens as a web page, 3,150 as Markdown. About 80% less to read, for the same words."

Careful: the bars are bytes, measured live on this laptop; Cloudflare's numbers are tokens, from their post. Don't put them on one scale or say "80%" about our page. If the tag says "Recorded", say "this is the reply I recorded when I built the deck", not "live".

My laptop · counted 4 Nov 2026

Of the 34 projects on my laptop,

2025scout Agent Reputation kleros-skills curatevision agent-on-trial Netlify upload Kleros x402 kit SDK 2025

32started this year.

21have a way in for agents.

13have no screen at all.

Way in for agents only 12
Both 9
Code library 1
Screen only 8
Documents 4

134 of my 142 commits were co-written with an agent.

17 Oct to 3 Nov · 8 others: syncs and merges

Counted 4 Nov 2026: project folders in ~/dev and ~/Desktop, dated by my first commit, sorted by what they ship; not this talk’s folder. Commits: all my repos, 17 Oct to 3 Nov, with a “Co-Authored-By: Claude” line. Only public repos named.

[1:00]

On entry: "That was then. Now: your agent reads the docs. Let me show you my own laptop. I counted it last night. Of the 34 projects on my laptop..."

Press. "...32 started this year." The two grey dashed tiles are from 2025, one of them is Scout.

Press. "21 have a way in for agents: a skill file, an llms.txt, a CLI, an MCP server or a paid API." The blue tiles have a screen too; the cyan ones are only for agents.

Press. "13 have no screen at all." The lit tiles: twelve made only for agents, plus one code library.

Press. "And 134 of my 142 commits from 17 October to yesterday were co-written with an agent. The other 8 are automatic syncs and merges. So I am not guessing about agent customers. I send them out every day."

If there is time: "Since July, every project of mine that is only a screen is for people in a room: a booth game, a check-in, a screen loop, talk decks."

Careful: say "34 projects", not 35. Tonight’s demo folder is not in the count; if you count it, the line becomes 35, 33, 22, 13, and you must not mix the two. "Which bucket a project falls in is my judgment call." "Co-written" is a floor: older commits made with an agent often have no trailer. Do not read out private repo names; only public ones carry a name on the tiles.

The way in

Agents come in through the docs.

What I see

What the data says

97%

of llms.txt files got no requests at all in May.

Ahrefs · about 38,000 sites

Claude Code was one of the top readers of llms.txt files, ahead of every AI search and assistant bot.

Today’s agent customers are builders’ agents.

Ahrefs, 15 Jun 2026: May 2026 requests to about 38,000 sites with a valid llms.txt · ahrefs.com/blog/llmstxt-study

[1:00]

On entry: "So where do agents come in? Not where we expected."

Press. "What I see at Kleros: most integrations come in through our docs, not our homepage. Builders point their coding agent at the docs and let it work it out."

Press. "The public data points the same way, with a twist. Ahrefs looked at about 38,000 sites with an llms.txt file, the file that tells an AI what a site is. In May, 97% of those files got no requests at all."

Press. "The ones that did get read: Claude Code was one of the top readers, ahead of every AI search and assistant bot."

Press. "So today’s agent customers are builders’ agents. A coding agent reading docs, on behalf of a person who is building something. Design for that one first."

Careful: no percentage for Kleros integrations (say no number), no company names. Say Claude Code was "one of the top" readers, not "number two" (Ahrefs’ own page is not consistent on its rank). The 97% is May 2026, about 38,000 sites.

Same content, different file

For agents, the format changes the result.

If we are humans, we do not care if the text is in a PDF, a doc or a PNG… for agents the difference is huge.

Me, on the Kleros live stream · 9 Sep 2026

Same words, three files

# Harbour Line: ferry tickets, for agents

Harbour Line runs return ferries from the Gateway of India, Mumbai. This is the page for software agents; people get the website at the same link.

## Routes and prices

Per person, return trip, in USDC. Children under 5 travel free.

| Route | Economy | Upper deck | Upper deck with guide |
|---|---|---|---|
| `gateway-elephanta` to Elephanta Island, 1 h | 2.60 | 3.20 | 4.80 |
| `gateway-mandwa` to Mandwa Jetty, 55 min | 3.10 | 4.20 | 5.60 |

Elephanta on Saturdays: out 09:00, 09:30, 10:00, 10:30, 11:30; back 12:30, 14:00, 15:30, 17:00. Boarding: `jetty-1` (stairs) or `jetty-4` (step-free). ID from age 18: `passport`, `aadhaar` or `driving-licence`.

## How to book as an agent

With the `ferry` CLI (`./bin/ferry` here; add `--json` for JSON):

1. `ferry departures --route gateway-elephanta --date 2026-11-07`
2. `ferry quote --route gateway-elephanta --date 2026-11-07 --depart 09:30 --return 14:00 --class upper-deck --board jetty-4 --passenger "Name,age,passport,P1234567" --assist "help boarding"` (one `--passenger` each)
3. `ferry book <quoteId>`: your human gets a yes card
4. `ferry wait <orderId>`: waits for the yes, pays with x402, prints the receipt. Exit 3 = still waiting, run it again.

With the API (JSON): `GET /ferry/api/departures?route=&date=` · `POST /ferry/api/quote` · `POST /ferry/api/orders {"quoteId","channel":"agent"}` · `POST /ferry/api/orders/<id>/pay` answers 402 with the price (x402 "exact", USDC on arbitrum-sepolia).

## Rules

- Ask your human before you pay. Every order waits for their yes, which only they can give, on the yes card.
- Prices are in USDC on a test network (Arbitrum Sepolia). No real money moves.
- Free cancellation until 18:00 the day before (Friday for Saturday).

More: [SKILL.md](/ferry/SKILL.md) · [llms.txt](/ferry/llms.txt)

Markdownharbour-line.md

The shop’s agent door, as plain text.

Gets the text itself: prices, times, commands.

`gateway-elephanta` … 2.60 | 3.20 | 4.80
… back 12:30, 14:00, 15:30, 17:00.
3. `ferry book <quoteId>`

Screenshotharbour-line.png

A screenshot of exactly that text.

Gets pixels. It has to look and read.
The commands are only a picture:

Scanned PDFharbour-line-scan.pdf

That screenshot, saved as a PDF, like a scan.

Gets the same picture, in a PDF wrapper:

/Subtype /Image /Width 1100 /Height 600

Markdown1,844 bytes

Screenshot136,331 bytes

74 times the size for the same words.

Quote: Kleros live stream, 9 Sep 2026, at 42:11, youtube.com/watch?v=gdtKyEL8EJM, lightly edited.
Measured on this laptop, 4 Nov 2026: the Harbour Line agent door as Markdown, as a screenshot,
and as that screenshot saved to PDF. Harbour Line is made up.

[1:00]

On entry: the quote. "Another thing I learned: for agents, the format changes the result. I said it on our live stream in September: if we are humans, we do not care if the text is in a PDF, a doc or a PNG. For agents the difference is huge."

Press: the three files. "Same words, three files. Our shop's agent door as Markdown, a screenshot of exactly that text, and that screenshot saved as a PDF, like a scan. To you and me, they look the same."

Press: what each one gives the agent. "Markdown gives the agent the text itself: the prices, the times, the commands it can run. The screenshot gives it pixels: it has to look and read, and the commands are only a picture. The PDF is the same picture in a wrapper."

Press: the size bar. "I tested it this morning: same question, same words, three files. A strong model got all three right on one short page. But the picture is 74 times the size, the commands in it can't be copied, and the longer and messier the document, the more that costs."

Careful: our test was one short page, three runs each; all three answered correctly (the Saturday return times and the upper-deck price), and times were noisy (Markdown 4 to 19 s, screenshot 5 to 80 s, scanned PDF 5 to 17 s). Do not claim a speed multiple like 5x. The 74x is file size, not tokens or time: the model's input was about the same, about 9,000 tokens, for all three. The PDF is 106,664 bytes, 58 times the Markdown. The quote is from a machine transcript of my own words, lightly edited.

The agent door · four pieces

What the agent should get.

01

Discover

A Markdown page and a skill file.

GET /ferry/
Accept: text/markdown
→ 200 index.md
  SKILL.md · llms.txt

02

Price

The price in the reply.

POST /quote
→ 200 quote
  12.80 USDC
  test network

03

Permission

A request the human can say yes to.

POST /orders
→ 202 awaiting_approval
  approveUrl
  /approve/?order=…

04

Pay

A receipt it can show.

POST /orders/…/pay
→ 402 Payment Required
X-PAYMENT: signed
→ 200 receipt

A CLI takes flags, not forms.

$ ferry quote --route gateway-elephanta --date 2026-11-07 --depart 09:30 --return 14:00 \
    --class upper-deck --board jetty-4 --passenger "Asha Rao,31,passport,P4417702" …

From this demo’s own shop (Harbour Line is made up). Payment: x402, HTTP status 402, test USDC on Arbitrum Sepolia.

[1:00]

On entry: "So what should the agent get from your door? Four pieces, in the order it needs them."

Press. "Discover. Ask for Markdown and you get Markdown: a short page and a skill file that teaches the steps. No cookie banner."

Press. "Price. The price comes back in the reply, before anything is booked: 12.80 USDC, on a test network. No form."

Press. "Permission. The agent does not pay yet. It gets a request its human can say yes to. We will see that card in a moment."

Press. "Pay. Once the yes comes, the agent asks to pay and gets HTTP 402, Payment Required. x402 finally uses it. It signs the payment and gets a receipt it can show: a reference, the transaction, a link to check it."

Press. "And one more: a CLI takes flags, not forms. Every field the website asks me to type is one flag here. Words are for the person. Flags are for the agent."

Careful: no numbers on this slide except the demo’s own price. 402 is a status code, not a statistic. The receipt is real only on the test network; never say "paid" before settlement.

Request from your agentTest network

1

Pay 12.80 USDC

to Harbour Lineapprox. 1,126 INR

TripGateway of India → Elephanta, return
WhenSat 7 Nov · out 09:30 · back 14:00
Seats4 × upper deck · 3.20 each
BoardJetty 4, step-free · help boarding

2

This matches your request9 of 9 points

Four people
Gateway of India to Elephanta
Saturday 7 November
Leave between 9 and 10
Back after 14:00
Upper deck if under 4 USDC
Step-free jetty
Help boarding for Sam
Ask before paying

3

Or choose insteadside by side

Cheaper10.40 USDCEconomy, same timesSaves 2.40 USDCChoose

Later12.80 USDCUpper deck at 10:301 hour later, same priceChoose

4

How to undo: free cancellation until Fri 18:00, refund to the same wallet.

5

What this yes covers: this order only. It expires in 10 minutes.

6

Not nowYes, pay 12.80 USDC

The yes card · on the phone

What the humanshould see.

1
A picture
The trip at a glance: the boat, the route, the island.
2
The total, and who gets it
Pay 12.80 USDC to Harbour Line. Then the order in four lines.
3
What it matches
One line per point you asked for. A difference gets a quiet dot.
4
Two options, side by side
Cheaper, and later. Not only the agent’s pick.
5
How to undo
Free cancellation until Friday 18:00, refund to the same wallet.
6
What this yes covers
This order only. It expires in 10 minutes.

Side by side changes the winner.

Two dictionaries: judged together, the bigger one won. One at a time, the one like new.

Agents took the first offer 60–100% of the time.

In a simulated market. So show the options.

Card: this demo’s yes card; Harbour Line is made up.
Hsee et al., Psychological Bulletin 125(5), 1999.
Microsoft, Magentic Marketplace, arXiv:2510.25779, section 5.4.

[1:15]

On entry: "Now the other side. The agent gets a whole door. The human gets one card, on the phone. What should be on it?"

Press. "A picture, so you know what this is before you read a word. Then the total and who gets it: pay 12.80 USDC to Harbour Line. Then the order in four lines."

Press. "What it matches: one line for every point of my request. A difference gets a quiet dot, not a red alarm."

Press. "Two alternatives, side by side: cheaper, and later. Not only the agent’s pick."

Press. "How to undo it, and what this yes covers: this order only, for ten minutes. A yes should never be a blank cheque."

Press. "Why side by side? An old finding: shown side by side, people paid more for a bigger dictionary with a torn cover; one at a time, they paid more for a smaller one in new condition. Side by side changes the winner. And in Microsoft’s simulated market, agents took the first offer they got 60 to 100% of the time. So show the options, not only the pick."

Careful: no dollar amounts for the dictionaries. Say "simulated market" for Microsoft, and 60 to 100% is from section 5.4 of the paper. The card on the slide is the demo’s; you will see the real one in the race.

The yes card · one per decision

The card is made for the decision.

Your agent was asked

“Book 4 ferry tickets, Gateway of India to Elephanta, Saturday, under 4 USDC each”

Sat 7 Nov09:30 → 14:00

Pay 12.80 USDC4 × 3.20

CheaperEconomy · 10.40

Later10:30, same price

Not nowYes, pay

Your agent was asked

“Order 40 team T-shirts in navy, sizes from the sign-up sheet”

Navy40 shirts

S8
M16
L12
XL4

XL is out of stock in navy. Grey instead, or wait 5 days?

Grey insteadWait 5 days

Your agent was asked

“Renew our design tool for next year”

Seats used, each month

You pay for 20 seats, you use 6.

Renew 20 seats2,400 USDC Switch to 10 seats1,200 USDC

Same agent. Same yes.A different card each time.

In Google Research’s test, people preferred a generated interface over a Markdown answer 82.8% of the time. Sites made by human experts still won.

Three made-up examples. Same agent, same kind of yes, a different card each time.
Google Research, “Generative UI: LLMs are Effective UI Generators”, arXiv:2604.09577, section 3.

[1:00]

On entry: the title. "That card can't be a fixed form. The agent knows what this decision needs, so it should make the card for it. Three examples."

Press: the ferry tickets. "Four ferry tickets. A trip, so the card shows the route and the times, the total, and two options side by side: cheaper, or later."

Press: the T-shirts. "Forty team T-shirts. Here the card shows the sizes, and the one question that needs you: XL is out of stock in navy. Grey, or wait five days?"

Press: the software plan. "Renewing a design tool. This card shows a chart, because the decision is about use: you pay for 20 seats and you use 6."

Press: the closing line and the study. "Same agent. Same yes. A different card each time. In Google Research's test, people preferred a generated interface over a Markdown answer 82.8% of the time. Sites made by human experts still won."

Careful: the three cases are made-up examples, with made-up prices and sizes. The 82.8% is Google Research's own test, not ours. Sites made by human experts still won, so do not say generated beats designed.

From my own work

Claude Code already does this. But the page sits next to the chat, not in it.

Claude Code

>

Show me three design directions for the deck

I made you a page with three options.
Open options.html

Next: the page is the answer

>

options.html

Three directions for the deckPick one

A · Quiet

B · Bright

C · Data

Save my pick

Disclaimer

In the cyberpunk version, none of this matters: we get a chip in the brain and the card goes straight in.
Until then, someone has to design the card.

Mock-up of my own workflow. The page and its options are examples.

[0:45]

On entry: the title rises. "This is already starting."

Press: the chat and the page appear, joined by a thin line. "When I work with Claude Code, it often answers with a page, not a paragraph: a set of options I can click, a chart, a questionnaire."

Press: the page steps back and an outline appears inside the chat. "But look where it lives: a file next to the chat. You open it, you come back. The next step is that the page is the answer, right in the conversation."

Press: the disclaimer, bottom right. "And yes, I know the cyberpunk version: a chip in the brain, and none of this matters. Until then, someone has to design the card."

Careful: this is my own experience as a user, not a product announcement; do not say Anthropic plans to build it. Do not name any company’s brain chip. Keep the chip line a joke, one sentence. The chat and the page on the slide are a mock-up.

A race.

One mission. Three ways in.

The mission, the same in every lane

4 ferry tickets, Harbour Line
Gateway of India to Elephanta
Sat 7 Nov, leave 9 to 10
Back after 14:00
Upper deck if under 4 USDC each
Step-free jetty
Help boarding for Sam
Ask me before paying

1

Agent · agent door

Claude Code with a CLI. Live.

Discover
Price
Permission
Pay

4 steps

2

Agent · human door

An agent on the website. Recorded.

Cookies
Account
Trip
Passengers
Review
Pay
Ticket

7 screens

3

Me · human door

Me, typing. Live.

Cookies
Account
Trip
Passengers
Review
Pay
Ticket

7 screens

A form nothing can autofill.

A cookie banner, an emailed sign-up code, a hand-made calendar, every passenger by hand, a card form.

Harbour Line is a made-up ferry company. Every name and ID number in the race is made up.

[0:45]

On entry: "A race." and the mission. "Let's test this. One mission: four ferry tickets to Elephanta this Saturday, with a few rules. Upper deck only if it is under 4 USDC each, the step-free jetty, help for Sam, who uses a cane. And: ask me before you pay."

Press: the three lanes. "Three ways in. Lane one: an agent, Claude Code, on the agent door: a Markdown page and a command line. Four steps. Lane two: an agent on the website, the door made for people. Recorded before the talk, real speed. Lane three: me, on the same website, typing live."

Press: the hard parts of the website light up. "And the website works the way most real ones do: a cookie banner, a sign-up code by email, a hand-made calendar, every passenger by hand. A form nothing can autofill. Let's go."

Careful: say once that Harbour Line and every name and ID in the race are made up. Lane 2 is a recording: do not call it live. The website is our own build, made slow on purpose to look like a real booking site; do not say it is a real operator's site.

Ready. One press starts the race.

Recorded race, real speed

The race

The live race board shows here.

The raceNot live here

Run the race from the laptop.

The race board needs the talk's own server. On the laptop, open http://127.0.0.1:4900/deck/ and come back to this slide.

1

Agentagent door

Reads the Markdown door, gets a price, asks me, pays. Live.

2

Agenthuman door

Clicks through the website made for people. Recorded.

3

Mehuman door

Me, typing on the same website. Live.

Harbour Line and every name in the race are made up.

[2:15]

Time box: 2:15 from entry. In the rehearsal the yes card came up about 30 seconds after the start, and the receipt about 5 seconds after the yes (longer if the test network is slow). Once lane 1 has its receipt, give your own lane about 20 more seconds, then stop typing and press on, even if lane 3 is not done: the next slide shows it as "Still going", and that is the point.

Before the talk: start the server, check the board shows three lanes saying "Ready" and lane 3 is on the cookie banner. Everything runs on this laptop: no phone, no terminal. On entry the slide clears a race left over from the rehearsal (finished more than six minutes ago, or started more than 20 minutes ago), so the board starts clean. A race you started in this talk is never cleared by going back and forth.

On entry: the board fills the screen, three lanes, all at zero. "Three lanes, same mission. Watch the clocks."

Press: the race starts. The deck tells the board, and the board tells the server. If the board does not start it within five seconds, the slide starts it on the server itself. Then click into lane 3 and fill in the website yourself. PageDown on the clicker still moves the deck while you type.

While you type, narrate lane 1 in short lines: "It read the Markdown door. It got a price. Now it is asking me." When the yes card appears in the phone frame on lane 1: "There it is: it stopped and asked me before paying." Read it out loud, the total and the two alternatives. In this demo the yes comes by itself after 7 seconds, the time it takes me to read the card. "Yes. Paid. With a receipt it can show."

Lane 2: "And this is an agent doing what I am doing, on the website made for me. Recorded before the talk, real speed." Keep typing; do not rush the form. The form being slow is the point.

If lane 1 fails or the model call hangs: press L. The board replays the last recorded race at real speed and the corner says "Recorded". Say "This is the rehearsal, recorded at" and the time. If the board itself is blank, go to the backup slide fd-race-backup.

Careful: the yes is simulated: the server says yes by itself 7 seconds after the agent asks. Say so if asked ("in a real shop that is my phone"). The website sign-in on my lane takes any name.

Careful: lane 2 is a recording, say so. Lane 1 pays on a test network (or the local test ledger, if that is what the receipt says), never real money. Do not claim speed numbers here; the next slide shows them. Everything named in the race is made up.

The race · Result Placeholder

The agent was done in seconds. Then it waited for me.

LaneTime, on one scaleTimeStepsTokens

1Agentagent door

–

2Agenthuman door

–

3Mehuman door

–

Placeholder: rehearsal numbers go hereNo race has been recorded on this laptop yet.

From the race recorded on this laptop: time, steps and tokens per lane. Harbour Line and every name in it are made up.

[0:45]

On entry: the headline and the three lanes, empty. The label top right says where the numbers come from: "Today's race" (live, from the laptop), "Recorded" with the date and time, or "Placeholder".

Press: the bars grow on one time scale and the numbers come in. "Here is the race we just ran. Lane one, the agent on the agent door. Lane two, an agent on the website. Lane three, me." Read lane one's time from the screen. Lanes 2 and 3 will usually say "Still going": say so, that is the point. Lane 2's line also says how long its whole recorded run took.

Press: lane one splits in two. "And look at lane one. This part is the agent working. This striped part is it waiting for my yes. It did not move until I said yes. That yes is the one thing I want to cross the narrow channel."

If the slide says Placeholder: "We will fill this in with the rehearsal numbers. You just saw the race live." Do not quote numbers from memory.

Careful: these are our own numbers from one race on one laptop, not a benchmark. Say "in our race" or "in our rehearsal, at" and the time on the label. Steps are counted differently per lane (commands and requests for the agent, clicks and fields for the website). If the agent's own work took more than a minute, the headline switches by itself to "The agent was done first"; say that instead of "in seconds".

Powerful models can use your website. That doesn’t mean they should.

Through the screenA robot in the driver’s seat.

53 steps 551k tokens about 17 min

and it finished only with a 14-step walkthrough

Through the APIA car built to drive itself.

8 calls 12k tokens 20 seconds

the same task

Reflex, Apr 2026 · Claude Sonnet, input tokens · screen: mean of 3 runs, with a walkthrough · API: mean of 5

[1:00]

On entry: the first line, and a robot draws itself into the driver's seat of a car made for people. "Here is what we just saw. Powerful models can use your website. They can click, scroll, type. Like a robot in the driver's seat: it works, the car was built for someone else."

Press: the second line, and a car with no driver's seat draws in on the right. "That doesn't mean they should. The other way is a car built to drive itself: a door made for the agent, with a price, a permission and a receipt."

Press: the numbers come in. "Reflex measured the same admin task both ways. Through the screen: 53 steps, 551 thousand tokens, about 17 minutes, and it only finished once someone wrote it a 14-step walkthrough. Through the API: 8 calls, 12 thousand tokens, 20 seconds." Pause. "If you are a builder, instead of a perfect UI you may want a CLI and an API."

Careful: do not say the screen agent "failed" next to the 53 steps: in the runs with those numbers it finished, with a walkthrough. With the plain request it did not finish, and Reflex gives no numbers for that run. 551k and 12k are input tokens, means of 3 and 5 runs, one model, one task, from Reflex's own blog. Do not turn it into a cost or a "times cheaper" number. Our race is not a benchmark either.

The dial · four notches

It does more. You see less.

What reaches you

You do every step

1Find and compare
2Create an account
3Type the code we sent
4Fill in four passengers
5Pay by card
6Keep the ticket

Before it pays

Pay 12.80 USDC

to Harbour Line · ferry for four

Yes, payNot now

Set once, before

Upper deck under 4 USDC a person

Ask me before paying more

After

The receipt: paid, four tickets

It answers. You do the rest.

Approve · a test case of mine

An AI juror said one thing, then did another.

It wrote“Refund buyer”

It votedPay seller

Check the action, not the explanation.

Set rules · my own code

In my shop prototype, permission is not yes or no. It has four settings.

NoneAgent attestationSigned delegationHuman confirmation

My newest tool only reads until you add --execute.

On its own · trust comes in steps

Watch closely
Check less
On its own

“Right now I don’t even check if my agent is doing correctly or not.”

A colleague who runs an AI juror, about 30 cases in

Source: Kleros Court V2 test network, court 8, dispute 76, 30 Aug 2026 · ruling “Pay seller”

Source: my own code · shop prototype, Aug 2026 · read-only tool, Nov 2026

Source: Kleros live stream, 26 Aug 2026 [1:15:53], lightly edited.

[1:45]

On entry, the dial sits on notch one, Ask. "And this is already starting: the agent acts on its own. Think of it as a dial with four notches. Notch one: Ask. The agent answers, and I still do every step myself. That whole column on the right is what reaches me."

Press: the dial turns to Approve, and the pane shrinks to one card. "Notch two: Approve. The agent does the work, and only one thing reaches me: a yes before it pays. That one card has to be right. A story from my own testing: an AI juror wrote a full verdict headed 'Refund buyer'. Then it cast the vote that pays the seller. So the yes card shows the action, not the explanation."

Press: Set rules, and the pane shrinks again to two lines. "Notch three: Set rules. I say once, up front, what it may do: upper deck if it's under four USDC a person, ask me before paying more. In my shop prototype, permission isn't yes or no, it has four settings: none, agent attestation, signed delegation, human confirmation. And my newest tool reads only, until you add --execute. Reading is free; acting is a choice."

Press: On its own, the pane is down to one line, the receipt. "Notch four: on its own. Trust comes in steps: watch closely, check less, then let it act. A colleague runs an AI juror. About 30 cases in, they said on our public stream that they don't even check its votes anymore. Their advice: watch the first five or ten cases closely, then check less often. That's where this goes. Then the only thing that still reaches me is what happened: the receipt."

Careful: the juror story is a test case on the test network, not a customer and not a real dispute; say "a test case of mine". Don't name the shop prototype's repo; it is a prototype, not a product. The trust steps are the public wording of the colleague's advice (watch the first five or ten cases closely). Their pronouns are not stated: say "they". Don't say the jurors are "Kleros's AI jurors": their owners run them.

Before · during · after

Three things cross the narrow channel.

Your agent’s side

Every departureEvery seat classThe price rules The timetableThe docsThe skill file The termsThe refund rulesYour mission The ID rulesStep-free accessSeats left The APICancellationYour budget

Your side

Before

BeforeYou tell your agent

Book 4 ferry tickets · Sat 7 Nov

Under 4 USDC eachAsk me before paying

During

On its own: no yes needed

DuringYour agent asks

Pay 12.80 USDC?

Not nowYes, pay

After

AfterYour agent reports

Paid 12.80 USDC

4 tickets · ref HL-EFQTJZ

If it goes wrong: a ruling in plain words

Cards: a recorded run of this talk’s demo,
on a test network. The ferry line is made up.

On its own, the middle one fades. The other two never do.

[1:00]

On entry: the picture from the start again. Left, dimmed, everything my agent takes in. Then the funnel, and on the right, my side, with three empty places. "So if less and less reaches me, what still has to cross that narrow channel? Three things."

Press: Before. The mission card goes from me into the funnel, and "Your mission" lights up on the agent's side. "Before: the mission and its fence. What to do, and how far it may go. Four ferry tickets for Saturday, under four USDC each, ask me before you pay. That one goes from me to the agent."

Press: During. The yes card comes back out of the funnel to me. "During: a yes, but only where I want one. One card, one decision, for this order only. That's the yes card you saw in the race."

Press: After. The receipt comes back the same way. "After: the receipt, and if something goes wrong, a ruling in plain words. A person reads it, not a log file. As a colleague put it on our stream, appeals go to people, and people don't read code."

Press: the yes card fades to an empty outline, and the last line appears. "Turn the dial to 'on its own' and the yes goes away. The other two never do. You always say what you want first, and you always get to see what happened after."

Careful: this is my model, not a standard; say "I think" if asked. Don't promise that the ruling exists for every purchase today: Kleros rules on disputes that are filed with it. The cards are from a recorded demo run on a test network, not a real purchase.

After the payment

If your agent gets it wrong, who decides?

With a Visa card, today

Reasons you can dispute a payment

Fraud

Not received

Not as described

Charged twice

My agent got it wrongno such reason

Visa Core Rules ch. 11: no reason for agent error

Under Visa’s rules, the cardholder answers for the agent’s purchase,

“as if the Cardholder initiated” it.

Visa Core Rules, April 2026 edition, §4.1.24.10

Kleros · a test we ran

We replayed 88 real disputes with AI jurors.

Typical case, people48.9 hours

Typical case, AI jurors9.8 minutes

72 of 88 rulings came out the same.

Kleros Court V2, Arbitrum One, Oct 2026 · our replay · typical = median of 83 paired cases

Being builtPeople stay as the expensive exit, on appeal.

[1:15]

On entry, only the headline. "The receipt always reaches me. So what happens when the receipt is wrong? My agent bought the wrong thing. Who decides what happens next?"

Press: the card list. "With a Visa card today, you dispute a payment for a reason: fraud, not received, not as described, charged twice. Visa's dispute rules have no reason for 'my agent got it wrong'."

Press: the Visa line. "And Visa's own rules say the cardholder answers for the agent's purchase 'as if the Cardholder initiated' it. So today, if your agent buys the wrong thing, you bought it."

Press: the Kleros beat. "This is the part I work on. We took 88 real disputes and replayed them with AI jurors. The typical case went from 48.9 hours to 9.8 minutes. And 72 of the 88 rulings came out the same as the people's."

Press: the last line. "That doesn't take people out. People stay as the expensive exit: an appeal will go to them. We're building that part now."

Careful: the four reasons are examples of everyday dispute reasons, not Visa's official list. If someone mentions American Express: Forkast reports Amex announced an Agent Purchase Protection; I am talking about Visa's dispute rules. The Visa quote was read in Visa's rulebook on 22 Oct (the PDF was blocked to scripts today). "Typical" means the median, over 83 paired cases; say "replayed", these were not live rulings. Don't say "customer". The human appeal path is being built, not live: never say "humans have overturned AI jurors", and never "a final ruling in minutes" (it's a first-round ruling, then an appeal window). Don't say "Kleros's AI jurors": their owners run them.

Leave knowing

Three things to take home.

1The agent door

What the agent should get

A Markdown page and a skill file
The price up front, then HTTP 402
A request you can say yes to
A receipt it can show

2The human door

What the human should see

A picture and the total
How it fits your request
Two options, side by side
How to undo, and what the yes covers

3After

Where trust is earned

The receipt, every time
Check the action, not the explanation
A ruling in plain words
People as the exit, on appeal

Up next · 15:40 · Agentic Hub

Harish Kotra

Can an agent buy from you?

Find, check, pay, get the result.

Send your agent

http://<laptop>:4900/talk

Runs on this laptop during the talk. Find me after to try it.

The agent’s channel is getting wider. The human’s is getting narrower.

Design what crosses the narrow one.

[0:30]

On entry: the hand-off. "Next, right here at 15:40: Harish Kotra, 'Can an agent buy from you?' That talk: find, check, pay, get the result. And if you want to try what I showed: after the talk, find me and point your agent at this address. Your agent gets Markdown, you get a page."

Press: the thesis, once more. "One line to take with you. The agent's channel is getting wider. The human's is getting narrower. Design what crosses the narrow one. Thank you."

Careful: check the room schedule for Harish's slot before the talk (the public listing has it at 15:40). The booth is Ethernet only, so the address works on this laptop's network only, while my laptop runs the server; if the address on screen still reads "<laptop>", read the LAN address the server printed at start, or skip this line.

Sources.

Every number in the talk, and where it comes from. Main links: /talk.md, under Sources.

Cloudflare: 16,180 tokens as HTML, 3,150 as Markdown
blog.cloudflare.com/markdown-for-agents · Feb 2026
Sizes of /talk and of the Harbour Line page
measured on this laptop · 4 Nov 2026
34 projects · 134 of 142 commits co-written
my own recount · commits 17 Oct to 3 Nov 2026
Claude Code, a top reader of llms.txt · 97% got no requests
ahrefs.com/blog/llmstxt-study · Jun 2026, May data
“For agents the difference is huge”
Kleros live stream · 9 Sep 2026 · at 42:11
The QR worked for people, not for agents
fortuvp.github.io/ask/condo and condo.txt · Oct 2026
Side by side changes the winner
Hsee and others · Psychological Bulletin · 1999
First offer picked 60–100% of the time, simulated
Microsoft, Magentic Marketplace · arXiv:2510.25779, 5.4
Race times, steps and tokens
our rehearsal recording · 4 Nov 2026
Screen: 53 steps, 551k tokens · API: 8 calls, 12k tokens
reflex.dev/blog, computer use vs APIs · Apr 2026
An AI juror wrote “Refund buyer”, voted “Pay seller”
Kleros V2 testnet · court 8, dispute 76 · Aug 2026
Four permission settings · read-only until --execute
my own prototypes · Aug and Nov 2026
A colleague no longer checks their juror’s votes
Kleros live stream · 26 Aug 2026 · at 1:15:53
Visa: no dispute reason for “my agent got it wrong”
Visa Core Rules, ch. 11 · read 22 Oct 2026
“As if the Cardholder initiated” it
Visa Core Rules 4.1.24.10 · April 2026 edition
88 disputes: 48.9 hours to 9.8 minutes · 72 of 88 matched
Kleros V2, Arbitrum One · court 34 replays · Oct 2026
People as the exit, on appeal: being built
KIP-93, forum.kleros.io · kleros-v2 PR #2592, draft

Backup · The race Placeholder

The race, recorded.

LaneTime, on one scaleTimeStepsTokens

1Agentagent door

–

2Agenthuman door

–

3Mehuman door

–

Placeholder: rehearsal numbers go hereNo race has been recorded on this laptop yet.

Lane 1, step by step

Discover–
Price–
Asks me–
My yes–
Paid–

Filled in from the race recorded on this laptop. Harbour Line and every name in it are made up.

[1:00]

Use this slide only if the live board fails (blank, frozen, or the model call never comes back). Everything is shown at once, no presses.

"The live board is not playing along, so here is the race from our rehearsal." Read the label top right: it says when it was recorded. "Lane one, the agent on the agent door: done in this time, and most of it was waiting for my yes. Lane two, an agent on the website. Lane three, me." Then the line at the bottom: "It found the door, got the price, asked me, and paid within seconds of my yes."

If it says Placeholder, there is no rehearsal recording on this laptop. Skip the numbers: "The point is the shape: the agent asks, then waits for a person." Go on to fd-race-result only if it shows numbers.

Careful: say "in our rehearsal" and give the time on the label; these are not today's live numbers. One race, one laptop, not a benchmark. If the payment line says local test ledger, say that, not "on chain".

Your next customer won’t click.

The agent’s channel is getting wider. The human’s is getting narrower.

One door, built for eyes.

Same link, two doors.

Of the 34 projects on my laptop,

Agents come in through the docs.

For agents, the format changes the result.

What the agent should get.

Discover

Price

Permission

Pay

What the humanshould see.

A picture

The total, and who gets it

What it matches

Two options, side by side

How to undo

What this yes covers

Side by side changes the winner.

Agents took the first offer 60–100% of the time.

The card is made for the decision.

Claude Code already does this. But the page sits next to the chat, not in it.

A race.

Run the race from the laptop.

The agent was done in seconds. Then it waited for me.

Powerful models can use your website. That doesn’t mean they should.

It does more. You see less.

Three things cross the narrow channel.

If your agent gets it wrong, who decides?

Three things to take home.

What the agent should get

What the human should see

Where trust is earned

Harish Kotra

Sources.

The race, recorded.

All slides

The live parts of this talk need the local server