flintspark Book a demo

Policy decides, not the model

In flintspark, policy decides what Flint, flintspark's agent, may do on each site and which actions need approval. The model never decides. An action whose risk is unknown counts as irreversible, and learning a new skill cannot loosen a policy.

  • Which routes Flint may useSet per site
  • Which requests it may replayA written allow-list
  • How fast it may goA rate limit per site
  • Whether it can lower a risk classNever. The model cannot

How are site policies set?

Each site has one policy, written by a person as configuration and reviewed like any other change. Flint does not write or edit a policy.

The platform shows a policy but does not change it. Every action Flint proposes passes the policy check before it is sent.

A sketch with illustrative values, not a screen from the product.

Which routes may Flint use?

A policy lists the execution routes, called tiers, that Flint may use on that site. There are four, fastest first.

  • An official API or an allow-listed page request, for reads by default.
  • A direct page address.
  • Recorded browser actions.
  • An AI agent reading the page, limited to the sites the workflow declares.
A sketch with illustrative values, not a screen from the product.

Which endpoints may Flint call?

Only the ones on the site's allow-list. An entry names the method, the URL pattern and, where needed, which part of the request body. A request that matches no entry is refused.

A POST counts as a read only when a person has listed that exact request as a read. Flint never infers it.

A sketch with illustrative values, not a screen from the product.

How are writes controlled?

Every action is classed as a read, a navigation, a reversible write or an irreversible one. If anything says irreversible, it is. An unknown action is classed irreversible.

The model can never lower a class. Flint's browser runs the same check on its own and acts on the higher of the two answers, so a server mistake cannot downgrade an action. By default an irreversible action waits for a person's approval. A write endpoint needs its own written entry, and still needs that approval.

  1. Read

    Looks at data. Needs no approval.

    Policy allows it
  2. Navigate

    Opens a page the policy allows.

    Policy allows it
  3. Reversible write

    Changes something that can be undone.

    Policy decides
  4. Irreversible

    Submit, send, pay or delete. Also any unknown.

    A person approves, by default

What limits does a policy set?

How fast Flint may go: a rate limit for each site, with a cap on how many pages it opens at once, enforced for every run that shares the browser. When a site answers with an error such as too many requests, Flint slows down.

Which fields are masked, which pages are blocked, and that exploring a site denies writes by default. Masked fields are hidden in the page before anything is recorded.

A sketch with illustrative values, not a screen from the product.

Can learning loosen a policy?

No. Learning adds skills and versions. A policy changes only when a person changes its configuration.

A workflow owner can switch on trust mode for one workflow. Policy then issues the approval grants, each still signed and single-use, and the switch goes into the audit log.

Questions about this page

Who writes a site's policy?

A person writes it as configuration, and it is reviewed like any other change. Flint does not write or edit policy, and the platform shows it read-only. A policy lists the routes Flint may use, the endpoint allow-list, rate limits, masked fields and a blocklist.

What happens when Flint is unsure whether an action is safe?

The action counts as irreversible. By default an irreversible action waits for a person's approval, and the model can never lower a risk class. Flint's browser runs the same check locally and acts on the higher of the two answers.

Can Flint learn its way around a policy?

No. Learning adds skills and versions, and it cannot allow a new route, add an endpoint to the allow-list or lower a risk class. A workflow owner can switch on trust mode for one workflow; policy then issues the grants, each still signed and single-use, and the switch is logged.

See Flint run

Bring one workflow you'd rather not do by hand. We'll show Flint learning one like it, and stopping before anything it can't undo.

We reply from hello@flintspark.ai.