Browse docs

Docs / start / use-cases

On this page

What you can build

Each shape below needs the same three things: a vocabulary the application shares with the Agent, material whose sources stay traceable, and writes that stay under policy. The differences are in the types, the relations that matter for retrieval, and which operations deserve to be Actions.

Use these as starting sketches. The declarations shown are illustrative, not a schema you must adopt.

Support assistant

An assistant that follows impact instead of guessing from ticket text.

Incident ── affects ──> Service
Incident ── owned-by ─> Team
Incident ── blocks ───> Release

Relations carry the work here. "Which open incidents affect checkout" is a traversal, not a similarity match, so the Selection Plan can show which links it followed and which it excluded. Projections keep responses small: summary returns title and severity, core adds impact and owner, full adds the timeline and evidence.

Actions worth declaring: acknowledge, reassign, escalate, resolve. Each needs a capability and a precondition, and resolve is a good candidate for approval.

Compliance review

A review that can cite the exact snapshot it read.

Controls, evidence, and exceptions are typed Units under a versioned release. Because snapshot identity binds model, corpus, release, and digest, a finding recorded in March can be reproduced in September against the same bytes. Licence and provenance metadata travel with the material, which matters when evidence comes from a third party.

Keep the review read-only at first. Granting an exception is the Action to declare later, with approval required and evidence appended.

Operations knowledge base

Runbooks where the risky steps are declared rather than described.

Procedures become Units with preconditions. The steps that change production become Actions with a side-effect class, an idempotency key, and a policy. An Agent can then ask for the rollback procedure and request the rollback, and the runtime decides whether that principal may run it. The Effect Plan comes back before anything executes, so a human sees what would change.

Data catalogue

Datasets, owners, lineage, and access rules retrieved through a profile you configure.

Lineage is a relation with direction and cardinality, so "what breaks if this table changes" is answerable. Visibility is enforced before candidate providers receive data, which is what keeps a private dataset from leaking through a relation or a score.

Design system guidance

The reference Frontend Design package publishes 797 Units of patterns, anti-patterns, rules, and examples that an Agent consults while writing interface code. It also ships a validator and an optional Skill, which makes it the most complete external package to read before you author your own.

Browse its Units in the registry or read the case study.

Choosing what becomes an Action

A Unit describes. An Action changes. When you are unsure, ask whether a wrong call needs an audit trail. If it does, declare it as an Action so it passes capability, policy, idempotency, and approval checks, and appends evidence. A Selection Plan can recommend an Action; it cannot grant the capability to run it.

Next

This page is rendered at build time from versioned Markdown in its owning repository.