The fastest route that works, step by step
An execution route is the way Flint, flintspark's agent, performs one step. There are four, fastest first: an API or allow-listed page endpoint, a direct page address, recorded browser actions, then an AI agent reading the page. A failing step drops one route, and only when that is safe to retry.
- 1 · APIAn official API, or an allow-listed page endpoint
- 2 · Page addressA direct address, with the details filled in
- 3 · Recorded actionsThe clicks recorded when the task was taught
- 4 · AI agentFlint reads the page and decides
When does Flint use an API?
First, wherever one exists. When a step reads data and an official API serves it, that is the route.
Where there is none, a request the page made during the demonstration can become the route, if the site's policy allows it. A check on the shape of the response guards it, and a change in that shape fails the step down to the next route.
Step · read the order list
- Route
- Tier 1 · replay of a read request
- Check
- The response has the expected shape
- If it fails
- Tier 2 · page address
Does tier 1 call a site's own endpoints?
Yes, for reads, and only from an allow-list. Each site's policy lists the requests Flint may replay, matched on method, address pattern and optionally the body. A person writes each entry. Flint never infers one, and any request that matches no entry is refused. Respect each site's terms when you add one.
A replay runs inside the logged-in browser profile, under the site's pace, carrying only the cookies and headers the page itself sent. A POST counts as a read only when the policy lists it as one. A write endpoint needs its own entry and still needs an approval.
Allow-list · supplier portal
- GET orders listread · listedallowed
- POST order searchread · listed with its body shapeallowed
- POST submit orderwrite · not listed as a readrefused
Read-only in the platform. It changes only through reviewed configuration.
What happens when a page changes?
Only the step that broke drops a route. Later steps start on their own routes.
If the step succeeds on a slower route, flintspark saves the fix as a new version of the skill, on probation. The next run tries it first, and it replaces the active version only after verified runs. The old version stays, so a person can roll back.
- Tier 1
The response shape changed. The step fails.
- Tier 2
The page no longer matches. The step fails.
- Tier 3
The recorded element is not found.
- Tier 4
The agent clicks the right element and the check passes.
When does a step not drop a route?
When it may already have acted. A step drops a route only if no action was attempted, or a check proved the action did not happen. Otherwise a lower route could do the same thing twice.
A step that may have acted goes to uncertain and asks a person. After a crash, a check runs before any irreversible step is redone. An expired login is not a failure either: the run pauses until a person logs in again.
After a failed attempt
- Nothing was attemptedor the check says it did not happendrop one route
- It may have actedan irreversible stepask a person
- The login expirednot a failed steppause
Why no captcha solving?
Because Flint works as itself. A captcha, a two-factor check or an expired session pauses the run and hands the browser to a person. Flint does not try to solve or get around it, and it does not disguise its browser.
The site's pace backs off while it waits. When the person resumes, Flint looks at the page again and retries the attempt on the route it was on.
Your turn · a check only a person should answer
The run waits. Nothing is retried around the check.
Questions about this page
Can Flint replay any request a site makes?
No. Tier 1 replays only requests on a per-site allow-list that a person wrote, matched on method, address pattern and optionally the body. Reads and exports are allowed by default; anything that matches no entry is refused. A write endpoint needs its own explicit entry and still needs an approval. Respect each site's terms when you add an entry.
What happens when a login expires during a run?
The run pauses and asks a person to log in again in the browser window Flint uses. An expired session is not counted as a failed step, so Flint does not drop a tier or write a repair. After the person resumes, Flint looks at the page again before its next action.
Does Flint slow down for the sites it works on?
Yes. Each site has a pace in its policy: a request rate, a short burst allowance, some random spacing and a limit on parallel tabs. When a site answers with a too-many-requests or unavailable error, or shows a challenge page, Flint backs off on that site.
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.