---
title: "Agentic Commerce Has an Octopus Problem | ZeroClick Blog"
url: https://zeroclick.ai/blog/agentic-commerce-has-an-octopus-problem
fetched_at: 2026-11-17T21:17:33.219Z
---

# ZeroClick API catalog and pricing

## API reference, endpoint lookup and prices

**[ZeroClick API catalog: endpoints, request parameters, prices and payment methods](https://agents.zeroclick.ai/zcj/wnvfgyyih6va/zeroclick/agent/guide).** Tell the seller advisor what you want to accomplish. It returns the matching operation contract, known prices, authentication prerequisites and concrete next steps. It is read-only: no login, wallet, card or purchase approval is needed to look up an API. It never executes a sale.

Available API services: Agent readiness report, Demo booking.

**Suggested next request:** GET [task guide](https://agents.zeroclick.ai/zcj/wnvfgyyih6va/zeroclick/agent/guide?goal=URL_ENCODED_TASK). Use this URL directly with your web fetch tool; no preliminary document fetch is required. Include known inputs and desired output in the goal. Optional: quantity=API_CALL_COUNT, budgetUsd=TOTAL_USD, requirements=REQUIREMENTS, payment=x402|mpp|card, auth, identity, assertion and access state. Omit unknowns. Do not equate output items with API calls or guess a call count.

For private task details or known request arguments, POST https://agents.zeroclick.ai/zcj/wnvfgyyih6va/zeroclick/agent/guide with Content-Type: application/json. Example: {"intent":"quote","goal":"YOUR_TASK","purchase":{"budgetUsd":"5.00"}}. Replace the task and budget with known facts; budgetUsd is a string. Include known buyerState, paymentCapabilities and arguments. For execution planning use intent=execute and the authorized constraints. Exact fields: [request schema](https://agents.zeroclick.ai/zcj/wnvfgyyih6va/zeroclick/agent/schema). GET and POST are alternatives, not mandatory stages.

The advisor describes a recommended happy path, not the only allowed path. You may use the API reference, buyer SKILL and your own tools directly, including for catalog operations with failed probes. Missing guidance or an unverified probe is not a policy refusal. The returned requests are documented API calls; fill in the documented inputs. Check usage cost AND minimum cash required against the budget. For needs_input, supply only missing facts. Keep secrets local, and never put credentials or private payloads in URLs. Unknown does not mean absent.

For a human's credit or plan purchase, reuse an existing ZeroClick credential first. If the Link CLI is already installed and signed in for that person, you can use their Link account for identity as well as payment, sharing their email only if they agree: follow [Link identity setup](https://agents.zeroclick.ai/zcj/wnvfgyyih6va/auth.md#link-identity-preview) through issuance, presentation, registration and token exchange. Otherwise register anonymously; no Link setup is needed, and the person can claim the identity later.

Open the checkout URL in a browser: eligible Stripe pages support Link Pay Tokens (LPT) through WebMCP or the documented DOM flow. Follow this storefront's /payment/card.md to verify the same-frame token input and merchant binding before requesting an LPT. Token submission requires payment authorization. If unavailable, privately send the checkout URL to your human to pay.

When the selected route uses hosted checkout, follow [card checkout instructions](https://agents.zeroclick.ai/zcj/wnvfgyyih6va/payment/card.md) before requesting a token. LPT goes only to Stripe, never to the advisor or ZeroClick API. Poll the returned statusUrl before using credit or retrying delivery.

Verify delivery; a payment or HTTP 202 is not completion. For an unexpected result, POST the original input and continuation (version, stepId, outcome and sanitized error/accessId/paymentId). Reconcile uncertain payments or writes before retrying. Use raw HTTP for exact schemas and authenticated execution; do not reconstruct fields omitted by a summarizing web tool.

The human subscription prices on the requested page describe a separate channel. Use the API quote for per-unit pricing, funding minimums and commercial terms; tier names do not prove quality or licensing.

For an alternative documented path, use the [buyer SKILL](https://agents.zeroclick.ai/zcj/wnvfgyyih6va/SKILL.md) or [full direct workflow reference](https://agents.zeroclick.ai/zcj/wnvfgyyih6va/llms-full.txt). Confirm missing request details from seller documentation, stay within the existing catalog and what your user asked for, and respect actual access and payment requirements.
Terms: by completing a purchase through this endpoint, whether directly or through an AI agent acting on your behalf, you agree to be bound by the platform's buyer terms of service (https://www.zeroclick.ai/legal/buyer-terms-of-service). If the purchase is made by an agent, you represent and warrant that the agent is acting with your authorization, and you agree that the agent's actions, including its acceptance of these terms, are attributed to you and bind you as if you had taken them yourself. Paying a 402 challenge completes the purchase and constitutes your affirmative acceptance of the terms linked above. The link travels in every payment challenge and receipt as `terms`.

# Agentic Commerce Has an Octopus Problem

**Zack Ewing** · Product · November 16, 2026 · 7 min read

[Blog](https://zeroclick.ai/blog)

---

The octopus is a fascinating creature, perhaps the best example of an alternative evolutionary path to sophisticated cognition on our planet (Peter Godfrey-Smith's *Other Minds* is the book to read on this). But unlike humans and the great apes, who evolved social structures for passing knowledge between generations, the entirety of an octopus's understanding is contained within a single lifetime. Most species are largely solitary and the mother dies as her eggs hatch, so almost no octopus ever meets its parent. They are born incredibly capable, work out how to optimize the tasks their lifecycle requires, and then everything they learned dies with them.

## Learn, then do

I have found it incredibly helpful in my product work in AI commerce to think of AI agents as entities with the same limitation. The moment you spin up an agent, it exists for the first time with its training data as a kind of genetics (everything it inherited from the humans who wrote the web, and nothing yet about the specific site in front of it), survives for the life of the problem it is solving, then dies with whatever it learned. So when you hand an agent a task you are waiting for it to both *learn* and then *do* what you asked, and for anything novel that gap shows up as a real difference in time-to-outcome.

For most of the ways people use agents today this limitation is not deeply felt, and I believe that is masking a deeper system problem. A retail sales analyst who asks their AI tool of choice to read a stack of spreadsheets and write an executive summary before a meeting in 5 minutes would have needed hours or days to do it themselves, so a directionally correct take in minutes feels like magic. The agent is still learning in real time, but when the alternative is a day of work a 5 minute answer is just as exciting as a 2 minute one.

## Commerce is where it gets felt first

As AI becomes our primary interface for more of the connected world, this "learn then do" behavior will start to create real friction, and I think commerce is where it gets felt first. At work an agent is competing against hours of human effort, so nobody notices the ten minutes it spent learning. When it is ordering your lunch it is competing against your own thumb, and the same ten minutes is the difference between a sale and no sale.

Consider a prompt that will be common in a few years. A high school teacher has offered their lunch break to a student who needs extra help, so they can't leave campus for the new restaurant down the street as planned, and they decide to order delivery during the 5 minute passing period between classes. Three ways this can go:

- In the first, we imagine the delivery apps as they exist today. The teacher opens one, sees the new restaurant is not listed, runs out of time choosing an alternative and is pulled away to the next class. They end up at the vending machine in the teachers' lounge, and neither they, the restaurant nor the app got their optimal outcome.
- In the second, our teacher is an early adopter of AI as a lifestyle assistant. They open their AI app of choice (Claude, ChatGPT, etc.) and ask it to "order me the chicken tikka masala from the new restaurant down the street for delivery here on campus at lunch." The agent finds the restaurant's own site and a couple of local listings, then gets stuck in a loop trying to work pages built for humans. At the next break the teacher finds it gave up after 10 minutes, and they end up at the vending machine again.
- In the third, the new restaurant's own site serves agents a version of its pages that states plainly what a buyer would need to know (the menu, whether it delivers to a given address, how to order and how to pay), which gives agents the generational knowledge sharing mechanism they never evolved. The agent skips the delivery apps, confirms the restaurant delivers to campus, orders the chicken tikka masala and opens a payment confirmation page inside the AI app. Every action in that loop still originates with the teacher's ask, the site just removed the ten minutes of trial and error between the ask and the order. Because the learning step is solved the whole loop takes under a minute, and a restaurant that is not on any delivery app gets a sale it would otherwise have missed.

## The environment is the teacher

Notice what changed between the second scenario and the third. The agent did not get any smarter and did not remember anything from a previous life, it simply landed somewhere that was built to be learned quickly.

This is the part of the octopus story I left out at the start. With no parent to learn from, an octopus has exactly two sources of knowledge, what it was born with and what it can work out from its surroundings, so the environment ends up doing the teaching. My favorite example, again from Godfrey-Smith, is Octopolis, a site off the east coast of Australia where gloomy octopuses, normally loners, live at unusually high density. It seems to have started with a single foot-long object half buried in the seabed that gave one octopus a den. That octopus ate scallops and left the shells, the shells piled up into a bed that later octopuses could dig their own dens into, and those octopuses left more shells. Nobody taught anybody anything, but each new arrival started somewhere earlier arrivals had already made easier. An agent is in the same position: everything it did not inherit through training it has to pick up from wherever it lands, and in commerce the place it lands is the seller's site.

## Won't the next model, or its memory, just fix this?

The natural objection is that the next model will simply navigate faster, and it will, but I don't think the learning step goes to zero, because what slows an agent down on a given site is that site's menu, inventory and checkout quirks, which will never be in a training set. That knowledge is held by the seller, so the seller is the only one who can hand it over. The biggest platforms are already doing exactly that (DoorDash and Instacart both launched ordering inside ChatGPT at the end of last year, and Stripe, Google, Visa and Mastercard are all building the payment rails). The restaurant down the street, the SaaS company with 40 customers and the API provider with a six page docs site cannot build a ChatGPT app, and that long tail is where the octopus problem lives, which is why the third scenario runs through the restaurant's own site.

The other objection is memory. Claude, ChatGPT and the rest now carry notes and skills across sessions, so a sharp reader will ask why the fix sits on the seller's side instead of inside the agent. I think the answer is that memory is one agent's notes from one visit, which may be months stale by the time they are used, while the seller's site is current and authoritative, and the seller is the only party with a reason to keep it right.

This is the future we are building toward at ZeroClick, and the belief underneath it is that tooling beyond the models themselves will serve a key role in enabling agentic commerce. So, going back to our octopus problem: humans escaped it by passing knowledge from parent to child and eventually by writing it down. Agent memory may eventually give agents something like a parent, but I am not convinced they need one. What they need is a better place to land, a site where everything the last visitor had to learn the hard way is already built into the surroundings. That is what we build for sellers: a version of your site that tells an arriving agent what you sell, what it costs and how to buy it, and the visibility to watch how agents actually do on your site so you know what to fix next.

Learn more at [zeroclick.ai](https://zeroclick.ai).
