Experiment 004: Autonomous Website Maintenance Loop
A bounded trial of agent-managed maintenance on static websites: source recovery, consistency fixes, build verification, and editorial lane definition.
Objective
Test how much routine static-site maintenance can be delegated to agents without lowering trust.
The focus of this trial was not visual redesign. It was operational stewardship:
- identifying the canonical source repo
- separating source from deploy artifacts
- making low-risk consistency fixes
- verifying builds after changes
- defining which editorial lanes are safe for autonomous work
Scope
This experiment used two live surfaces:
hellointernet.lolfullyautomated.enterprises
The work stayed inside a bounded maintenance lane:
- source-tree inspection
- content and footer cleanup
- public-link cleanup
- planning and operating-note creation
- build verification
System Configuration
- local agent operator with repo and shell access
- static Astro sites with multiple surfaces/subdomains
- build-first verification loop
- review-gated treatment for design changes, public claims, and deployment wiring
Work Performed
Source-of-truth recovery
The largest operational risk was not broken code. It was source confusion.
A canonical source repo for hellointernet.lol had been displaced by an artifact-like checkout, while a more current source clone existed in a temporary path. The maintenance loop first had to identify the real editable source tree, move it into the durable projects location, preserve the artifact checkout as backup, and rebuild all surfaces from the corrected location.
Consistency and credibility fixes
Low-risk changes were then applied:
- normalized footer wording across
hellointernet.lolsurfaces - removed placeholder social/contact links from
fullyautomated.enterprises - wrote site-management and redesign planning notes
- audited weak or placeholder surfaces such as
linksandnow
Verification
All touched Astro surfaces were rebuilt after changes. Static redirect/utility surfaces were also checked for expected files.
Observations
What worked
A narrowly scoped maintenance loop works well when the task is concrete and verifiable.
Agents were effective at:
- locating source-vs-deploy confusion
- performing repetitive consistency edits
- writing planning artifacts
- rebuilding surfaces to catch breakage
- separating safe autonomous work from review-gated work
What failed or remained constrained
Agents should not invent freshness.
Pages like now become low-trust quickly if there is no explicit source of truth behind them. The same applies to research claims, product status, and current-focus language.
The experiment also showed that visual redesign is a different class of task from maintenance. It needs direction and approval, even when the implementation work is agent-capable.
Unexpected behavior
The most consequential failure mode was administrative rather than technical: a previous agent had cloned or worked from a temporary source location, which made it easy to confuse a deploy-shaped tree for the canonical repo. Once that ambiguity existed, every later edit became higher-risk until the source tree was re-established.
Human Intervention Required
Human input was still required for:
- deciding acceptable public language
- approving homepage direction changes
- defining what counts as canonical current-state information
- setting the boundary between demo scaffolding and real published work
Current conclusion
Autonomous website maintenance is viable in a bounded lane.
It is strongest when the work is:
- reversible
- verifiable
- local to known repos
- low-risk in public claims
It becomes much less reliable when the task depends on:
- ambiguous source-of-truth inputs
- subjective design decisions
- current-state claims without a maintained upstream source
- synthetic content standing in for real work
Artifacts produced
This trial produced:
- source-repo canonicalization for
hellointernet.lol - consistency fixes across multiple surfaces
- homepage redesign brief for
hellointernet.lol - surface-copy audit for
hellointernet.lol - content-transition and seed-work planning notes for
fullyautomated.enterprises
Next iteration
Next iterations should test a slightly more ambitious lane:
- real
linksupkeep - a source-backed
nowworkflow - publishing real lab entries as experiments are completed
- formal analyst/operator handoffs between Greg and Rick