flintspark Book a demo

Skills flintspark learns once and reuses

In flintspark, a skill is a versioned unit of work that Flint, flintspark's agent, learns once and reuses across workflows. A new skill is trusted only after two verified runs. When one step breaks, the fix is kept as a new version on probation, and an earlier version stays available to roll back to.

  • Read the weekly regional pricesAdmitted · used by 2 workflows
  • Read next year's forecastAdmitted · repair on probation
  • Fill and submit the supplier requestCandidate · 1 verified run
A sketch with the energy demo's names, not a screen from the product.

What is a skill?

One thing Flint can do on a site, such as "get the weekly price series". It is what a teach session produces, after a person reviews it.

Each skill has a typed interface, which says what goes in and what comes out, and one implementation for each site it works on. Each step of an implementation can run on several routes, from an API down to an AI agent reading the page.

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

How does Flint reuse a skill?

Before it learns anything new, Flint looks in the skill graph: the connected map of every skill learned, the sites each one works on and the workflows that use it.

  • If a chain of skills it already has covers a task, Flint drafts a workflow from them instead of a new skill.
  • A normal run is offered admitted skills only. Candidates show up only while someone is teaching.
A sketch, not a screen from the product.

When is a new skill trusted?

After two verified runs. A run is verified when its checks passed and the result was confirmed by a person or by a deterministic check. Until then the skill is a candidate.

One exception exists. An operator can confirm a single live run of a just-taught candidate before its two verifications. That confirmation goes into the audit log, and every irreversible step in the run still needs its own approval.

A sketch, not a screen from the product.

How are skills versioned?

A change to a skill is never an edit in place. Each repair or improvement is written as a new version that points to the one it came from.

The version in use stays admitted while a newer one is tried. When the newer one is admitted, the earlier one is kept, so a person can roll back. A run already in progress keeps the version it started with.

A sketch, not a screen from the product.

What is a probation version?

When one step breaks and a slower route fixes it, only that step is repaired. The fix is saved as a new version on probation, a single-step repair, instead of a rewrite of the whole workflow.

The next run tries the probation version first and falls back to the admitted one if it fails. After two verified runs it takes over. A newer repair replaces it and rejects it.

A session that has expired is not a failure, so it never produces a repair.

A sketch, not a screen from the product.

Questions about this page

What is a probation version of a skill?

When one step breaks and a slower route fixes it, flintspark saves the fix as a new version of the skill on probation. The next run tries it first and falls back to the admitted version if it fails. After two verified runs it takes over, and the earlier version is kept so a person can roll back.

Does Flint use skills that are not yet trusted?

Not in a normal run. Retrieval returns admitted skills only, and a candidate is a newly learned skill that has not passed two verified runs. One exception exists: an operator can confirm a single live run of a just-taught candidate, which is logged, and every irreversible step in it still needs its own approval.

Can a skill be rolled back?

Yes. Every change to a skill is a new version that points to the one it came from, and the earlier admitted version is kept. Runs already in progress keep the version they started with.

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.