Self-Healing Website
One context layer, many agents. The website notices its own defects, proposes the fix, and waits for a human to promote it.
Detect defects
Diagnose root cause
Propose a fix
Await promotion
Report on itself
↓Every capability reads the same context layer — that is what keeps a fleet of agents consistent instead of merely parallel.
Shared context layer
Change cadence
Agent memory
- This run's working set — What the agent has already looked at in this pass, what it has ruled out, and the evidence it is holding for the proposal it is about to write.
Hive memory
- What other agents learned — Fixes that were accepted, fixes that were rejected and why, and the patterns other agents on the same context layer have already confirmed.
Operational state
- Live site signals — Current crawl and accessibility findings, page performance, error rates, publish queue, and which pages changed in the last hour.
Organizational knowledge
- How Orlando Health writes and works — Brand and editorial standards, accessibility commitments, the review and approval path, and who owns which section of the site.
Ontology
- What the entities are — Services, locations, providers, and the relationships between them — the vocabulary every agent on this context layer has to agree on.
↓Systems are how the layers get their facts; the classification of the system sets the ceiling on the use case's risk tier.
Systems of record and action
- Builder.ioPublicReads page structure; writes proposed changes as unpublished drafts.
- VercelCompanyBuild and preview environment the proposal is rendered into for review.
- Marketing Data HubCompanyResolves services, locations, and providers to their canonical records.
- SiteImproveCompanySource of accessibility, broken-link, and quality findings.
- Google AnalyticsSensitiveBehavioral signal used to rank which defects matter.
- WorkfrontSensitiveWhere a promotion request becomes a tracked work item with a named approver.
Governance
Tier: MediumA2 — Human-reviewed, then published or sent broadlyThe rubric suggests High for the systems attached and the autonomy declared. The stored tier is Medium.
- Why this tier
- Medium in its first form: every change the agent proposes is reviewed by a named editor before it reaches a live page, so the blast radius of a bad inference is a rejected suggestion. The tier rises as soon as promotion becomes automatic.
- The ladder
- Rung 1 — detect and report only. Rung 2 — propose a diff against a page, human promotes. Rung 3 — auto-promote a narrow, enumerated class of fixes (broken links, alt text, stale hours) with post-hoc review. Each rung is a separate approval, not a gradual drift.
- Never
- No clinical content, no provider credentials, no pricing or coverage language, and no page that collects patient information. Those pages are read-only to the agent at every rung.
- Audit
- Every proposal, promotion, and rejection is logged with the agent's evidence and the reviewer's name. The log is the artifact a regulator or an internal audit reads, so it is written before the fix ships, not after.
↓Nothing reaches a live page without passing through a promotion the agent cannot perform itself.
Built by increment
Rung 1 ships alone and is useful alone. Nothing later in the ladder is a prerequisite for value now.
Portable by construction
The context layer is not website-shaped. The next agent that needs services, locations, and editorial standards inherits it.
Living
The layers have different half-lives, which is why they are drawn separately. Treating ontology like operational state is how context rots.