flintspark Book a demo

AI agent glossary

This glossary defines the terms flintspark uses for AI agents, each in one or two plain sentences: teach by demonstration, skill, verified run, execution route, approval grant, human in the loop, audit trail and redaction. Each entry says how flintspark, whose agent is Flint, uses the term, and where on this site to read more.

What is teach by demonstration?

A way to build an AI agent's skill by doing the task once while it is recorded, instead of writing code or a script. In flintspark, a person shows Flint, flintspark's agent, a browser task once; Flint drafts a skill from the recording, and a person reviews it before it is used.

How flintspark uses it. It is how every flintspark skill is learned today; teaching a task by describing it in words is not proven yet.

How a person teaches Flint

What is skill?

In flintspark, a skill is a reusable, versioned unit of work that Flint, flintspark's agent, can perform on a site, such as "get the weekly price series". Each skill has a typed interface and per-site implementations, and each step can run on several routes.

How flintspark uses it. Flint reuses an existing skill before it learns a new one, and a repaired step is kept as a new version of the skill.

How a demonstration becomes a skill

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

What is verified run?

In flintspark, a verified run is a run whose checks passed and whose result was confirmed by a person or a deterministic check. A new skill is admitted after two verified runs.

How flintspark uses it. A skill that has not passed its verified runs is only a candidate, and normal workflows do not use it.

How a skill is verified

A sketch, not a screen from the product.

What is execution route?

In flintspark, the execution route is the way Flint, flintspark's agent, performs a step. There are four tiers, fastest first: an official API or an allow-listed page endpoint (reads by default); a direct page address; recorded browser actions; an AI agent reading the page. A failing step drops one tier only when it is safe to retry; a step that may already have taken an irreversible action asks a person instead.

How flintspark uses it. Flint runs each step on the fastest route that works, and a failing step moves down one route only when retrying is safe.

Which route each step takes

What is approval grant?

In flintspark, an approval grant is a signed, single-use permission for one exact irreversible action. Flint, flintspark's agent, cannot take an irreversible action without a valid grant, which a person gives by default or policy gives in trust mode. A grant cannot be reused or replayed.

How flintspark uses it. The browser side checks each grant's signature before it acts, and a workflow owner can switch on trust mode for one workflow, which is logged.

What needs a person's approval

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

What is human in the loop?

A design where people answer questions and approve key actions inside an automated workflow. In flintspark, by default people approve every irreversible action, answer questions Flint, flintspark's agent, cannot, and can take over at any time.

How flintspark uses it. Product release, disposition and signing are never run by Flint, in any mode; Flint hands them back to a person.

How people watch and steer Flint

What is audit trail?

In flintspark, the audit log is a hash-chained record of approvals, grant mints and checks, take-overs, trust-mode switches, redaction drops and purges, where any later change to a past entry is detectable. Run steps and evidence are kept separately as run records.

How flintspark uses it. Only approvals and safety events are hash-chained; run records are kept separately, for 90 days by default, and the log is not a validated GxP audit trail.

How the audit log is protected

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

What is redaction?

In flintspark, redaction is removing listed personal data (known values, emails, phone numbers, card numbers, addresses and per-site fields) before anything is stored, shown in the platform or sent to a model. It fails closed.

How flintspark uses it. If redaction cannot complete, the data is dropped rather than kept; text inside images is not detected yet.

How personal data is handled

What is not defined here yet?

This month's glossary holds eight terms. The rest of flintspark's vocabulary is on the way, and four longer guides will go deeper on the largest ideas. They are not published yet, so they are named here without links.

Last updated

  1. Teach by demonstrationBuilding an agent's skill by doing a task once while it is recorded
  2. Human in the loopHow people answer questions and approve key actions inside an automated workflow
  3. Audit trail for AI agentsWhat to record, what tamper-evident means and what it does not prove
  4. Computer-use agentsHow an AI that operates software through its screen works, and where it fails

Questions about this page

What is the difference between a skill and a workflow?

In flintspark, a skill is a reusable, versioned unit of work that Flint, flintspark's agent, can perform on a site, such as getting a weekly price series. A workflow is a sequence of skills, checks, questions and approvals that delivers one business outcome. One skill can be used by many workflows.

Does a verified run mean a person checked it?

Not always. A verified run is a run whose checks passed and whose result was confirmed by a person or by a deterministic check. flintspark admits a new skill after two verified runs; until then it is a candidate and normal workflows do not use it.

Why do the entries say "by default" about approvals?

Because a workflow owner can switch on trust mode for one workflow. Policy then issues the approval grants instead of a person, each grant is still signed and single-use, and the switch is logged. Product release, disposition and signing are never run by Flint, in any mode.

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.