Why AI Agents Still Won't Pay for Anything — Jan Curn, Apify

Jan Curn· Apify20:37

Read the talk

Why AI Agents Still Won't Pay for Anything

The title slide states the talk’s central question.Open full source frame
The title slide states the talk’s central question.

Jan Čurn follows agent payments from the need for a budget to discovery and settlement, using Apify’s experience with x402 to show where the friction remains.

From a talk by Jan Curn

At a glance

Ideas worth remembering

  • Existing accounts and API keys already let agents use paid services; autonomous payment becomes useful when an agent needs to discover and buy a new service during a task.

  • A funded wallet does not establish whether an unfamiliar seller will deliver the promised service at the promised price. Trusted relationships and directories are potential sources of that signal.

  • For variable-cost jobs, a seller must handle the interval between authorizing payment and settling the final charge. Apify’s upfront-charge-and-refund workaround shifted trust to the buyer and incurred an extra transaction.

  • Čurn reports that batch settlement addressed Apify’s double-spending concern through an escrow-like arrangement, but the talk does not detail its implementation.

A crowded field of payment protocols

Agents already use paid tools, but Jan Čurn sees little autonomous discovery and purchasing. That gap matters to the Apify founder: at the time of the talk, its marketplace had about 80,000 community-created tools, called Actors, for web data extraction and automation. Čurn says Apify paid their creators almost $2 million in the preceding month. Autonomous purchasing would give those creators another way to reach customers. 0:12

The Apify marketplace slide illustrates the Actors discussed in the opening section.Open full source frame
The Apify marketplace slide illustrates the Actors discussed in the opening section.

Longer autonomous tasks may require an agent to spend without stopping for approval at every purchase. Payment companies have rushed toward that possibility: Čurn names Coinbase’s x402, Google’s AP2, OpenAI and Stripe’s ACP, Visa’s TAP, Google and Shopify’s UCP, and Stripe’s MPP among a much longer list. The companies have reason to compete for transactions on their existing networks; the number of protocols alone says little about how much agents are spending through them.

Čurn sees a particular opening for crypto, despite describing himself as no crypto enthusiast. His case rests on small transaction costs and the absence of card-style chargebacks: a merchant can choose to issue a refund, but a buyer cannot simply reverse a payment in the way he describes for card transactions. He worries that automated abuse could make chargebacks especially costly at agent scale. These are arguments about the payment properties, not evidence that any particular network is already widely used.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

The volume gap and the payment funnel

How much money was actually moving? Čurn presents figures from x402 and MPP transaction-monitoring sites: approximately $1 million on x402 and $1,500 on MPP over the preceding 30 days. The comparison concerns the blockchain transactions those sites monitored at that moment; it is not a measure of all spending by agents or all possible payment channels. 7:14

A shortage of services does not seem to him to explain the gap: payment directories already list many offerings. Čurn estimates that Apify and several other listed providers together earn far more than $1 million a month from agents using their services, principally through existing accounts and API keys. He proposes a funnel to explain why autonomous purchases are rarer: a builder must first need them, set up payments, give the agent a way to discover trustworthy services, and finally let it buy.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:44 · section reference included

The existing competitor is an API key

For web data, a developer can create an Apify account, put its API key in a cloud environment or connect through MCP, and pay the resulting usage on a credit card. The agent can call the services the developer has already chosen. For this workflow, autonomous payments compete with a convenient prearranged account. 9:43

The reason to add a spending mechanism is a different capability: letting an agent find a service it was not given in advance and purchase it during a task. Čurn expects better models and longer-running personal agents to increase that need. He suggests building trust incrementally, perhaps giving an agent $5 for a job, then $50, then $500 as it proves useful. He considers a bounded monetary loss easier to tolerate than some failures already possible when an agent has write access to a system or sensitive data.

The economics of model use also affect the choice between building and buying. Čurn claims that subsidized token access can make it seem cheap for an agent to construct a solution from scratch. If that subsidy diminishes, buying an existing service may become more attractive. Both the size and duration of such subsidies, and their effect on purchase decisions, are his estimates and forecasts.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:14 · section reference included

A wallet still needs a trustworthy seller

Even a builder who wants autonomous payments has setup work to do. Crypto terminology and wallet creation can be unfamiliar, and giving an agent a wallet takes more effort than paying a familiar service by card. Čurn’s proposed simplification is for an agent platform to integrate payments into its existing account: a user might allocate $50 of account credits for service purchases. He presents this as a possible future integration, not a feature established by the talk.

A funded agent then faces a separate question: will a service at an unfamiliar URL deliver the promised result at the advertised price? Finding a weather API says nothing by itself about its reliability. Čurn compares the missing signal to asking a friend for a barber recommendation: trust can travel through an existing relationship.

Directories and marketplaces can provide another signal by acting as intermediaries. Čurn points to reviewed connectors in Claude’s directory, ChatGPT apps, and Apify’s marketplace as examples of places that can help users find services. His proposed paths are a network of known, trusted services or a directory operated by someone the buyer trusts. Neither a listing nor the ability to pay guarantees that every individual transaction will succeed.

12:1212:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:12 · section reference included

From an HTTP 402 response to settled payment

The final step is billing for the work itself. A fixed price per API call can fit a weather forecast, but a compute job or another agentic job using language models may consume an amount of resources unknown at the start. Those jobs need variable, pay-per-use pricing. Čurn says x402 added batch settlement and MPP added sessions to address such billing.

How does the basic x402 payment path work? The client sends an API request and receives an HTTP 402 response with payment requirements. It signs a payment and repeats the request; the server checks it with a facilitator. The facilitator then submits the transaction to the blockchain and reports settlement after confirmation. That sequence separates checking the payment from knowing that money has actually transferred. 17:12

When to do the work is a business decision. If a server starts an expensive compute job before settlement and the payment fails, it can bear the cost without receiving money. Waiting for settlement before performing work reduces that exposure, at the cost of putting confirmation on the path to delivery.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:12 · section reference included

Pay-per-use exposed a double-spending problem

Apify’s experience shows why variable pricing took more than replacing a fixed price with a maximum. Čurn describes three approaches in x402’s progression:

  • exact: The initial scheme charged a fixed amount. Apify could use it to collect money before starting, but it did not directly express an unknown final bill.
  • upto: This allowed a variable charge, but Čurn says it did not prevent an agent from spending the same funds elsewhere before later blockchain settlement. That left Apify exposed while usage accumulated.
  • Batch settlement: Čurn says the later scheme solved that issue by holding funds on-chain in an escrow-like arrangement, then allowing charges for usage. He does not explain the underlying implementation in detail.

Before batch settlement, Apify worked around the problem with exact and a refund. In Čurn’s example, it charges $10 upfront; if the job consumes $2, it returns the remaining $8 when the work finishes. This protects the seller from waiting to collect, but the buyer must trust the seller to refund the balance. It also requires two blockchain transactions, adding fees. Čurn says Apify added batch settlement only shortly before the talk. 19:12

Čurn expects the payment funnel to become practical as these pieces improve and encourages service providers to investigate selling to agents now. His own example supplies the qualification: accepting a protocol can involve decisions about settlement timing, variable usage, counterparty trust, and transaction costs. The infrastructure is progressing; substantial autonomous payment volume remains his forecast.

Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

18:12 · section reference included

Resources

  • Announcing MCP connectorsArticle

    Čurn explains how Apify Actors can use authenticated external MCP services, including how its proxy handles credentials and permissions.

Transcript download
Full transcript

Loading transcript…