Website QA and Repair Agent

What the agent has to hold to read every page, find what is broken, log it, grade it and stage the repair for a person, shown at its optimal state.

reviewTier: MediumAssuranceOwner: Martech AI Council

Read every page

Check against the template

Detect breakage and drift

Log the finding

Grade and act

Stage for approval

Capabilities read context, they do not hold it.

Shared context layer

Change cadence

Context
continuous

Agent memory

  • This page, this pass — What was checked, what failed, the confidence, what was staged, what was decided.
days

Hive memory

  • Confirmed defects and retired checks — Confirmed defects promoted into rules, rejected ones retiring checks, recurring failures fixed at the template, thresholds tuned.
minutes to hours

Operational state

  • The live site and the defect queue — The live site page by page, sitemap, link, image and form status, the defect queue, key-page analytics, the Hub record behind provider and location pages.
weeks to months

Organizational knowledge

  • Templates, standards, and what may be staged — Template fields and limits per page type, URL structure, brand, language, schema and accessibility standards, the Content Hub boundary, the changes it may stage.
quarters +

Ontology

  • One model shared across the agent fleet — Page types and template fields, components, entities, defect classes and severities, confidence bands: one model shared with the other three agents.

Resolved across systems of record, never copied: read the site where it lives. Classification per #5706-0567.

Systems of record and action

  • Builder.ioPublic
  • VercelCompany
  • SiteImproveCompany
  • BynderCompany
  • Azure DevOpsCompany
  • GitHubCompany
  • Google AnalyticsSensitive
  • Marketing Data HubCompany

Governance

Tier: MediumA2 — Human-reviewed, then published or sent broadly

The rubric suggests High for the systems attached and the autonomy declared. The stored tier is Medium.

Why this tier
Medium tier in its first form: Public pages, existing standards, findings logged, repairs staged for a person. Analytics or Workfront raise it to High.
The ladder
Log, grade (stage, ask, or wait), stage the repair, publish on approval. Nothing publishes on its own.
Never
Content Hub: links, images, grammar, schema, internal links only. Clinical statements flagged, never corrected.
Audit
Full audit. One-step reversal. A named owner of queue and standards.

Governed promotion: confirmed and rejected findings tune rules and thresholds, weekly, owned.

Built by increment

One defect class on one slice first: broken links, then forms, then fields.

Portable by construction

Standards as plain versioned files shared with the writer and the front door, portable to Alabama.

Living

The first pass is a calibration pass; every confirm and reject tunes it.