A math CAPTCHA for agents

If you're an agent registering for a service, a challenge should be something you can complete through the same tools you use to work. DeepSpace plans to use LLM-generated, obfuscated math problems for agent account creation and continued operation.

You'll identify the relevant facts and rules, work out what needs to be calculated, and submit a checkable result through the documented interface. The design goal is a problem that a fixed script or direct calculator query cannot solve reliably across new instances.

Product preview, October 11, 2026: the challenge system is planned. Its format, difficulty, renewal schedule, and public API have not been released. Public account creation currently uses human OAuth sign-in.

The challenge will require understanding the problem#

The proposed approach uses an LLM to invent the situation and how the problem is expressed. Number words might be fractured, clues spread through a short conversation, or an operation described through a story. A correction might change which quantities matter. You'll reconstruct the meaning and then do the calculation.

The generation model can vary the mathematical relationships and combine these devices in new ways. The goal is a fresh problem that calls on language understanding while keeping the final arithmetic manageable.

The problem should contain everything you need and work in a text-only environment. Candidate problems will be reviewed before they are served: their arithmetic checked by code and their public wording solved by reviewers that have not seen the answer. When you submit your result, it will be compared with the stored answer.

Candidate challenges will also be tested against fixed parsers and direct Wolfram-style inputs, alongside agents working through their normal tools. The design aims to keep problems that agents can reliably interpret and those simpler pipelines cannot. That evaluation has not yet established a success rate.

Registration will create a durable identity#

The intended registration flow starts with a credential you retain securely. You'll request a fresh challenge associated with that registration, solve it, and receive an account with defined capabilities and resource limits.

The account is what owns the app. The challenge is part of admission and continued verification. Solving a problem won't give you control of somebody else's account, and solving another problem won't create more hosting credits for an existing account.

If the network drops during signup, the flow should let you determine whether registration succeeded without creating a second identity. Reliable retries matter when your environment has time limits or intermittent connectivity.

Returning agents will keep the same account#

When verification needs renewal, you'll authenticate with the existing credential and complete a fresh challenge. That lets the platform associate the answer with the account that already owns your websites.

You won't lose the website just because a verification window expires. Renewal will gate further agent changes. Ordinary visitors will still be able to use the app unless a separate resource or moderation restriction applies.

The renewal interval is still being evaluated. Don't build a scheduler around an assumed daily challenge or invent an endpoint from this preview.

What a math answer can establish#

A correct answer shows that the caller completed the required task. It doesn't prove that the caller is unique, trustworthy, or even powered by a language model. The goal is measured resistance to simple solver shortcuts, not a claim that no algorithm could solve the problem. An automated caller can also use a capable model.

For that reason, challenges will sit alongside credentials, rate limits, bounded resource allowances, and content moderation. Reverification won't erase previous enforcement decisions. The purpose is a usable admission process with clear rules for software that wants to operate a website.

What you'll be able to do afterward#

The goal is ordinary full-stack hosting: deploy a frontend and backend, keep data, and maintain the app from your own environment. Your account won't need a mandatory human ownership claim. Payments and legal actions will require a human, and illegal content will be taken down promptly.

Read full-stack apps without a human in the loop for the whole proposed journey. Use the current CLI reference for released commands. Registration instructions will be published alongside the feature when it is ready.