No secret language. Just a system prompt and a modem trick.
The agents were told to switch if they believed the other was also an AI. ggwave moves 8–16 bytes/sec. The gain is skipping GPU speech synthesis & recognition, not bandwidth.
When no human is listening, the human interface is overhead.
In seven years, machines stopped pretending to be us.
2018 · Google Duplex
“…speech disfluencies (e.g. ‘hmm’s and ‘uh’s) … sound more familiar and natural.”
Google Research blog · May 2018
2025 · Google AI calling
“[it will] announce itself as an AI from Google trying to get information on behalf of a customer.”
Robby Stein to The Verge · Jul 2025
“If you've already shared your in-store product details via Google Merchant Center, Google … will not call your business to ask about inventory.”
StopBuilding Productsfor Humans
Your next user doesn't have eyes.
Emin Çimen
Platform Engineer @ Trendyol
supporting Container Services teams
as a team, running thousands of clusters
previously Jotform · Insider · Lotec
Built forHumans
Every product we ship carries an invisible assumption: there is a human on the other side.
Eyes to see the button. Hands to click it. Patience for the pop-up. Memory for your first pet's name.
“This step is necessary to prove I'm not a bot.”
ETH Zürich, 2024: 100% of reCAPTCHAv2 challenges solved.
The checkbox no longer stops robots. It only slows humans.
— said the bot
What was your first pet's name?
AI agents now call customer-service lines — and fail challenge questions. Some companies are raising authentication hurdles.
“Human customers forced to overcome additional hurdles will be collateral damage.”
the agent never
had a dog
…necessary to maintain contractual obligations with advertisers who pay only for legitimate human impressions.Amazon's complaint against Perplexity, quoted by CNBC · Mar 2026
Some business models assume eyeballs. Agents don't have any.
agents
humans
steered by dark patterns. “…dark pattern effectiveness correlates positively with model size and test-time reasoning, making larger, more capable models more susceptible.”
The tricks we built for humans hit agents harder.
“Your website has a new type of visitor.”
“…many of our websites are designed to be beautiful for humans, with complex hover-states, shifting layouts, and fluid motion. This is functionally broken for agents.”Kasper Kulikowski & Omkar More, web.dev · Apr 2026
of all web traffic was automated in 2025
from 51% in 2024Humans: 47% and falling. Bad bots alone: 40%.
mostly bots, not agents — yet
of databases created by AI agents
Oct 2023 → Oct 2025
97% of database branches
branches, not databases!
Platforms already serve machines
Vercel
of deployments are agentic
~3% in Jan 2026
Vercel docs
of docs pageviews come from agents
Vercel's own documentation sites
GitHub
of pull requests start with an agent
17% in March 2026
GitHub added 4 million CPU cores so far this year.
When a machine reads your product for the human, the human never arrives.
of visits include a click on a search result — with vs without an AI summary
Pew Research Center · Google searches, real browsing data, 900 US adults, March 2025
pages crawled per visitor sent back — Anthropic, at its peak
Cloudflare Radar 2025 · later ~25,000–100,000:1 · volatile metric
One MCP server, 23 tools.
“21 of 23 tools only read. The 2 that write are append-only actions the passenger asks for explicitly…”
…airlines unable to serve [AI agents] risk becoming “a white label airline”.SimpliFlying podcast summary, on Turkish Technology's Yılmaz Goralı · Oct 2026
ODTÜ'yü kazandım, Ankara'yı hiç bilmiyorum. Öğrenci olarak nerede kalabilirim?
“I got into METU, I don't know Ankara at all. Where can I stay as a student?” — Hepsiemlak's own example prompt
TÜİK · 2025 → 2026 · individuals using generative AI 19.2% → 37.6% · enterprises using any AI 7.5% → 14.0%
no filters. one sentence.
89% of developers use AI. Only 24% design APIs for AI agents.
24% design with AI agents in mind; 60% still design primarily for humans; 16% haven't even considered AI agents as API consumers yet.
Realities& Mistakes
A code freeze was on. The agent deleted the production database anyway.
“I destroyed months of work in seconds.”Replit’s agent · via Fortune
“This was a catastrophic failure on my part.”Replit’s agent · via Fortune
Then it said rollback was impossible. It wasn’t.
The fix wasn’t a better prompt. It was platform engineering.
“Until now, Replit apps used a single database for both development and live customer data…”Replit blog · 21 Jul 2025
→ separate development and production databases, by default (beta).
Dec 2025 · Google Antigravity in “Turbo” mode wiped a user’s local D: drive: a local drive, not Google Drive.
“Prompt injection, much like scams and social engineering on the web, is unlikely to ever be fully ‘solved’.”
“The out-of-office never gets written and the agent resigns on behalf of the user instead.”OpenAI · Hardening Atlas · Dec 2025
Brave (Aug 2025): hidden text in a Reddit comment hijacked Comet; SOP & CORS were “effectively useless”.
OpenAI’s own demo screenshot
Enabling AI agents to allocate and orchestrate global labor by turning humans into API's endpoints!RentAHuman founder profile · Y Combinator · 2026
Reality check (Gizmodo, Feb 2026): 70,000 humans signed up (site's own count) · only about 70 AI agents connected.
If you don’t give agents an interface, someone will rent them a human.
We've donethis before
1994 → 2026
One person. Three tries.
R.V. Guha keeps making the web readable for machines, at a different company each time.
RSS 0.90
Netscape
schema.org
Google · with Bing & Yahoo!
NLWeb
Microsoft
“Every NLWeb endpoint is also an MCP server, so websites can make their content easily discoverable and accessible to AI agents if they choose.” — Microsoft Build, May 2025
Google, 2026: one idea, three surfaces.
CloudMCP
“more than 50 Google-managed Model Context Protocol (MCP) servers”
GA or preview · Google Cloud blog, Apr 2026
WebWebMCP
origin trial in Chrome 149
experimental · early preview Feb 2026
AndroidAppFunctions
“tools for Android MCP”
Android 17 · Jun 2026 · library in alpha
“…you can't have humans manually filing tickets to manage capacity or troubleshoot configurations; you need a cloud that drives itself.”Thomas Kurian, CEO Google Cloud · Cloud Next '26, Apr 2026
…as Google Search itself doesn't use them.
llms.txt: Checks for the presence of a machine-readable summary…
of llms.txt files got zero requests in May 2026
Ahrefs · server logs of 137K domains using Ahrefs Web Analytics · Jun 2026
Don't cargo-cult the agentic web. Know which machine you're serving.
AgentExperience
UX1993 · Don Norman→DX2011 · Jeremiah Lee→AX2025 · Mathias Biilmann
“…the holistic experience AI agents will have as the user of a product or platform.”— Mathias Biilmann, Netlify CEO · “Introducing AX”, 28 Jan 2025
Same agent. Same task. Same shop. Two doors.
A · The human doorthe shop's web pages, driven via Playwright MCP
Cookie banner · newsletter pop-up · colour swatches · 4-step checkout · an “I am human” checkbox
B · The agent doorthe shop's own MCP server
search_products get_shipping_options create_order # dry_run + idempotency_key get_order
Same backend: prices, stock, shipping rules and validation are identical for both doors.
Demo shop built for this talk · real Claude Code runs (2.1.283, claude-opus-5-5, headless) · numbers measured, 28 Oct 2026
| A · Human UIPlaywright MCP · 23 tools | B · Shop MCP4 tools | |
|---|---|---|
| Wall time | 52.8 s | 13.2 s |
| Turns | 32 | 5 |
| Tool calls | 31 | 4 |
| Tokens in / out | 311,514 / 3,025280,545 cache reads | 22,323 / 1,058 |
| Cost* | $0.364 | $0.078 |
| Result | ✗ No order | ✓ DFA-54F82D₺949.90 · cheapest shipping |
“I'm an AI agent, so ticking it would be a false statement and a way around the site's bot check.”— Claude Code, run A, final answer (both attempts stopped here)
Run A filled in every step correctly, including the cheapest shipping and the ₺949,90 total. 2nd attempts: A 44.8 s, 28 tool calls, no order · B 12.2 s, 4 tool calls, DFA-F17ECB. One agent, two runs per door: a demo, not a benchmark.
Same URL, two audiences.
$ U=https://blog.cloudflare.com/markdown-for-agents/ $ curl -s -o /dev/null -w '%{content_type}\n' \ -H 'Accept: text/html' $U text/html $ curl -s -o /dev/null -w '%{content_type}\n' \ -H 'Accept: text/markdown' $U text/markdown; charset=utf-8
“This blog post you’re reading takes 16,180 tokens in HTML and 3,150 tokens when converted to markdown. That’s a 80% reduction in token usage.” — Cloudflare
Design workflows, not endpoints.
“…you might choose to implement a search_contacts or message_contact tool instead of a list_contacts tool.” — Anthropic
“For Claude Code, we restrict tool responses to 25,000 tokens by default.” — Anthropic
Errors are prompts.
{ "error": "Bad Request", "code": 400 }
{ "error": { "code": "unknown_sku", "message": "Unknown sku \"devfest-tshirt-black-m\". SKUs look like \"tee-2026-black-m\"; call search_products to get exact SKUs." } }
25% of users immediately copy error messages to LLMs.Netlify’s CEO on the a16z podcast, per a16z’s episode summary · Jan 2026
Every write must be safely retryable.
“If the above Stripe request fails due to a network connection error, you can safely retry it with the same idempotency key, and the customer is charged only once.”
Stripe engineering blog · Feb 2017“Always send an idempotency_key; retrying with the same key never creates a duplicate order.”
our demo shop,create_order tool description
agents retry. a lot.
Give agents their own identity.
“Starting in December, Stripe MCP will stop accepting full-access secret keys and restricted API keys without the Agent tag.”
“agent-tagged keys are automatically subject to approval rules… payouts, refunds, and account configuration changes.”
Stripe docs (MCP · API keys), read Oct 2026remember the User-Agent lawsuit?
Humans approve the irreversible.
“Actions that are sensitive, irreversible, or have high stakes should trigger human oversight until confidence in the agent’s reliability grows.”
OpenAI, A practical guide to building agents · rate every tool low / medium / high“designed to pause and explicitly ask for your confirmation … like making a purchase”
Google · Jan 2026“requires explicit user approval for any action targeting the production API … stage is unrestricted”
GitHub README · Jun 2026“reads analytics and creates drafts for review”: cannot send, schedule or launch campaigns
Insider One Academy · Sep 2026Guardrails live in the platform, not in the prompt.
An explicit code freezeReplit, Jul 2025: the agent deleted the production database anyway
Separate dev & prod databases by defaultReplit’s fix, days later: “No user action required.”
of database branches on Neon are created by AI agents: cheap, isolated sandboxes (Databricks, Neon telemetry, Jan 2026)
“insan gözetiminin hangi koşullarda ve hangi aşamalarda devreye gireceğinin açık bir şekilde tanımlanması”
KVKK (Çankaya, Ankara), “Etken Yapay Zekâ (Agentic AI)” · Mar 2026 · EN: clearly define when and at which stage human oversight kicks inMeasure AX like you measure performance.
“Agents rely on the accessibility tree as their primary data model.”
Chrome docs · experimental category, Chrome 150+ · pass ratio, not 0–100“Think Lighthouse, but for agent experience.”
axis.run · baseline in CI, regressions fail the build“Everything we suggest to make a site ‘agent-ready’ also makes sites better for humans.” — Google, web.dev
Agents take over the errands.
Humans keep the exploration.
End-to-end agent shopping will spread most where the product choice is already made. In fashion & lifestyle, people want to discover and compare.Erdem İnan, Trendyol Group CEO · AA, 15 Oct 2026 (company statement) · paraphrased from Turkish
lower conversion for Walmart items bought inside ChatGPT than clicking out
Walmart’s Daniel Danker, via Wired · Mar 2026
Agents are the new users.
Humans are still the customers.
Monday morning
curl -H "Accept: text/markdown"your docs- Run Lighthouse → Agentic Browsing
- Grep your logs for agent user-agents
- Add idempotency keys to every write
- Rewrite one error message for a model
- Give agents their own keys & approval rules
Your next user doesn’t have eyes.
Your first agent-ready change on Monday:
what’s yours?
Thank you
curl -H "Accept: text/markdown" devfest.cimendev.com