How Skills Work

Skills are produced and used entirely behind the scenes. This page explains the lifecycle: how a Skill is learned from a Simulation, how versions accumulate, and how the agent replays a Skill in a later run.

Distillation: learning a Skill

When a simulation task completes successfully, Marketrix reviews the steps the agent took during that run. It looks for multi-step patterns the agent repeated or could have performed as a single unit — for example:

Filling out and submitting a form
Signing in
Searching and then selecting a result
Stepping through a multi-screen wizard

Each pattern it finds becomes a Skill: a name, a one-line description, and the sequence of actions that make it up. One-off operations — a single click or a single navigation — are deliberately left out, because those are already part of the agent’s built-in primitives and there’s nothing reusable to capture.

Distillation only runs when a simulation task meets all of these conditions:

The task completed successfully (failed or stopped runs are not distilled).
The run produced enough steps to be worth capturing — by default, at least three.

Distillation is best-effort and runs after the simulation finishes. If it doesn’t produce a Skill, the simulation’s own result is unaffected.

Versioning

Skills are versioned and immutable. Marketrix never edits a learned Skill in place. When the agent learns a refined version of a workflow it already knows, it records a new version of that Skill rather than overwriting the previous one.

This means every Skill carries a full version history — its lineage — so you can see how a workflow evolved over time and which version is currently active. The most recent, non-deprecated version is the one the agent uses.

How the agent replays a Skill

In a later run, an active Skill is offered to the agent as a single named action alongside its built-in primitives. When the agent chooses to use a Skill, Marketrix replays the captured sequence of steps as one unit.

Some Skills accept inputs — for example, a search Skill might take the search term, or a login Skill might take credentials. The agent supplies these values when it invokes the Skill, and sensitive values (such as credentials) are filled in securely at execution time and are never logged.

Scope and availability

Per Application — A learned Skill belongs to the Application it was learned in, and is only offered to agents working on that same Application.
Built-in primitives vs. learned Skills — The agent always has its baseline browser primitives available. Learned Skills are layered on top of those.
Experience level — How many of an Application’s learned Skills an agent draws on can vary with the agent’s experience level: a less experienced agent leans on the baseline primitives, while a more experienced agent has the full learned-Skill catalogue available.

Skills Overview What Skills are and why they matter.

Using Skills Review, inspect, and deprecate learned Skills.

Simulations The runs Skills are distilled from.