Yes! We let agents discover and pay for tools through one prepaid balance and we support recovering stored paid results after a connection drops, without paying for the call again. There are limits: a one-hour recovery window and an 8 MB response cap.

I think the interface and payment rail can evolve independently. An API can charge a prepaid balance or a card, while a browser agent can use a wallet.

The agent experience includes what happens after a timeout. Being able to recover a paid result without buying it again could matter as much as making the first payment easy.

The more useful chart would show independent buyers coming back to pay for the same service again. Much harder to manufacture than two wallets keeping each other busy.

Digital services have a smaller version of the same gap. A payment can succeed while the API returns unusable data. The agent still needs a way to check that it got what the task required.

Those permissions absolutely need to be enforced where the payment is signed. Allowed recipients, budgeting and a clear approval threshold give the agent room to work without making the prompt the only spending control.

For the newsletter failure, did the alert distinguish ‘nothing sent’ from ‘send status unknown’? That changes whether the next step should be a retry or a delivery check.

A useful recovery test: the tool succeeds, but its response never reaches the agent. Can the resumed run recover the result without repeating the action or paying twice?

Model choice becomes part of the workflow once it affects both quality and cost. Dot is Astra based and does not count to spending limits, which is pretty neat

Missing credentials are especially frustrating when the only evidence is buried in a transcript. An access request that names the blocked operation, required scope and owner gives the human something they can resolve immediately.

What drove the 40x gap: tool definitions in context, extra model turns, or execution costs? Did discovery happen once per run or on every step?

Ask your agent: “Find recent X posts about our product and link to the ones worth reading.” TaskFuel gives it access to full post text and engagement data through Otto, without a separate X developer account. taskfuel.ai/blog/x-search-fo…

Sales, trading and content, all from one agent paying per call. This is the agent economy in practice ⚡

What's the reason for their rate limiting you think? Why throttle usage of your product :)

The missing price without JavaScript is especially revealing. I’d add a check that asks an agent to identify the price and complete the intended action from a fresh session with no JS.

I’d separate routine token refresh from asking the person to sign in again. Would you require fresh approval when permissions or spending authority change, even if the two-week window hasn’t expired?

Maybe some kind of agentic voting, with publicly accesible scores showing if the results were appropiate could partially solve this?

I wonder how many shops, especially big ones, will force users to KYC to buy with their crypto wallets. Privacy is another concern to tackle

Would be interesting to see an answer from your agent when asked why it did so many changes compared to the human