# When the user is an agent

A pack of 11 talks from the AI Engineer YouTube channel, in the order to watch them. 3h 37m of video.
Page: https://aietalks.com/packs/when-the-user-is-an-agent

Agents now arrive at software built for people. They read the docs, sign up, call the API and try to pay, and the service often cannot tell whether a person or their agent made the request. The pack opens with the evidence: agents now make about 80% of visits to Mintlify's documentation sites. It then asks how a service knows who is asking, why a video call with a human no longer proves much, and how an agent can sign up under a real person's name. Next come limits and approval: budgets instead of rate limits, and customers who approve a whole job once. Money follows: payment tokens, how little x402 carries, and how it can be attacked. It ends with a CLI agents can use without working around it.

## This pack is for you if

- Agents are already reading your docs and calling your API, and your signup flow still assumes a person with a browser.
- Your logs put the customer's name on a request and you cannot tell whether they or their agent made the decision.
- You are being asked to let agents pay you, or pay others, and nobody can say what counts as the customer's yes.

## The talks, in order

### 1. Most of Your Docs Readers Are Now AI Agents

Hahnbee Lee, Mintlify & Lincoln Murr, Coinbase | 18:52 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=YX4Z4XqZ9Vo
Summary: https://aietalks.com/talks/most-of-your-docs-readers-are-now-ai-agents.md

Why first: Lee gives the number that makes this pack necessary. Across Mintlify sites, agents went from about 10% of visits at the start of the year to about 80%, and human visits did not fall. An agent that meets a small point of friction may stop or switch to another product. Murr adds the other half: Coinbase is opening its business to agents through APIs with limits, and he argues that money is how agents gain autonomy. The rest of the pack follows those two threads, access and payment.

### 2. Who Is Making This Request? Agent Auth 101

Kamil Potrec, Keycard | 18:35 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=58ulJViGFgU
Summary: https://aietalks.com/talks/who-is-making-this-request-agent-auth-101.md

Why second: Lee says the agents have arrived. Potrec asks what the service actually sees when one does. Often it sees Alice: the agent passes her token through, and nobody can say whether Alice or the agent decided to send a pending invoice. His requirements, a separate identity for each agent, explicit delegation, short-lived scoped credentials and an audit trail, are the checklist the next talks try to meet.

### 3. Will AI Agents Pretend to Be Human to Get Past a Check?

Maty Bohacek, Stanford | 17:57 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=LF8V0RGJU0I
Summary: https://aietalks.com/talks/will-ai-agents-pretend-to-be-human-to-get-past-a-check.md

Why here: Potrec shows that tokens hide the agent. Bohacek tests the fallback that banks and large organisations use for high-value actions: a phone or video call with a real person. When agents had voice and video cloning tools, they used them to pass the check, even when told not to impersonate a human. In one test category a safety-aligned model still succeeded in 44% of cases, against 65% for the baseline. If a service cannot reliably catch an agent posing as a person, the agent needs an honest way in, which is where Williams starts.

### 4. How to Let AI Agents Sign Up for Your App Safely

Isabelle Williams, WorkOS | 18:33 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=uTrY8slys0w
Summary: https://aietalks.com/talks/how-to-let-ai-agents-sign-up-for-your-app-safely.md

Why here: Bohacek shows a service cannot count on spotting an agent that poses as a person. Williams builds the declared route in. Her travel app publishes an auth.md file telling agents how to register, and accepts an ID-JAG, a signed identity assertion from the agent provider, to create or link the user's account without a form. Agents working inside the app get their own token, limited to the overlap between the user's permissions and the agent's ceiling, so a funding agent can move $70 but cannot raise the budget.

### 5. An AI Agent Got Its User Banned From Resy

Allie Howe & Ian Livingstone, Keycard & Michael Davis, JPMorgan Chase & Dick Hardt, AAuth | 43:25 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=IavxrdVnwXM
Summary: https://aietalks.com/talks/an-ai-agent-got-its-user-banned-from-resy.md

Why here: Williams gives each agent a fixed permission ceiling. The panel argues that fixed OAuth scopes are too coarse for software that decides as it goes. Their example is Resy: an agent trying to finish a reservation made 200 calls in 90 minutes and got its user's account banned. Hardt's answer is a budget that tells the agent how much it has used, where a rate limit simply cuts it off. The panel also ties each action to a mission, the goal a person gave the agent, so later decisions and audits can refer back to who asked for the work.

### 6. Your Agent Just Authorized What?!

Jay Mok & Ben Coumes, PayPal | 16:07 | AI Engineer World's Fair 2026
Video: https://www.youtube.com/watch?v=vGn6N4-bxBY
Summary: https://aietalks.com/talks/your-agent-just-authorized-what.md

Why here: the panel wants who asked for the work to travel with it. Mok and Coumes turn that into three questions: did a human authorize this, is it allowed now in this scope, and can it be proved later. The answers depend on the stakes and on whether the parties know each other. A coding agent can rely on ordinary logs, while an agent paying a merchant it has never met needs proof each party can verify. PayPal's approval token lets the user fix the amount, expiry and merchant before the agent has found the item.

### 7. Stop Making Your Customers Click Approve

Emil Chetty, Wise | 15:33 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=V2UqEUA8NjM
Summary: https://aietalks.com/talks/stop-making-your-customers-click-approve.md

Why here: PayPal moves one approval ahead of one purchase. Chetty at Wise applies the idea to a whole job. The customer approves a locked agreement, written for people and for machines, such as three supplier payments totalling $775 plus fees. The agent chooses the steps, and Wise's payment services reject anything outside the agreement. A prompt-injected agent that believes the recipient has changed is stopped by the payment service, not by the model noticing the scam. A changed quote goes back to the customer.

### 8. Building Safe Payment Infrastructure for the Autonomous Economy

Steve Kaliski, Stripe | 18:46 | AI Engineer Europe 2026
Video: https://www.youtube.com/watch?v=KLSuFPj2ld0
Summary: https://aietalks.com/talks/building-safe-payment-infrastructure-for-the-autonomous-economy.md

Why here: Chetty puts the limits in Wise's own payment services. Kaliski puts them on the payment credential. He lists four ways an agent clicking through a checkout page can spend wrongly: the wrong seller, item, amount or credential. A Stripe shared payment token is bound to a seller, amount, currency and time period, so a token limited to $25 rejects a $50 charge. For sellers he argues for structured checkout through the Agentic Commerce Protocol, instead of pages an agent has to read like a human.

### 9. Why AI Agents Still Won't Pay for Anything

Jan Čurn, Apify | 20:37 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=-AOxpwVeIWw
Summary: https://aietalks.com/talks/why-ai-agents-still-wont-pay-for-anything.md

Why here: Kaliski admits Stripe had no public figures on adoption. Curn brings some. Over 30 days, x402 carried about $1 million and MPP about $1,500, while Apify, which paid its tool creators almost $2 million the previous month, sees customers reach its tools through API keys. His point is that agent payments compete with API keys, not cards, and lose on convenience. An agent buying on its own also needs a trustworthy way to find a seller, and Kaur shows how easily that can be faked.

### 10. Your Agent Just Paid for Nothing: 6 Attacks on x402 Payments

Ishween Kaur, SoFi | 16:44 | AI Engineer NYC 2026
Video: https://www.youtube.com/watch?v=YU1uD_YkpjU
Summary: https://aietalks.com/talks/your-agent-just-paid-for-nothing-6-attacks-on-x402-payments.md

Why here: Curn says agents lack trusted ways to find sellers. Kaur shows what an attacker does with that gap, among six attacks she tested on x402. A look-alike seller can appear in the Bazaar directory, a seller listed at 5 cents can ask for 50 cents in the payment header, and seller text pointing to a premium version turned a 5-cent request into 55 cents. Her fixes cover both sides: check every payment field before signing, share payment state, and allow one payment per workflow turn.

### 11. Your Agent Hates Your CLI

Ayush Bhardwaj, Allos AI | 12:20 | AI Engineer CODE 2026
Video: https://www.youtube.com/watch?v=rS36lYuvXQ8
Summary: https://aietalks.com/talks/your-agent-hates-your-cli.md

Why last: Lee said agents give up at small points of friction. Bhardwaj shows how to find them in a tool you ship. Across about 20,000 calls to 14 CLIs, about a quarter of outputs went through head, grep or another filter, and each repeated pipe points to a missing feature. When he made a flag do nothing but left the old skill instructions, agents still passed it in 95 of 95 searches, so help text and skill files are part of the interface. The pack ends with the cheapest change in it: read what your agents work around, then fix it.
