Owner and release gate
James
Sets the offer, judges whether the work is credible and controls every external action.
Multi-agent delivery system
Site Revamp: several AI agents, one accountable result.
Site Revamp used specialised AI agents, shared project context, independent quality checks and explicit human approval gates to turn market research into 45 public local-business website demos.
It is proof that AI can support substantial delivery work without being allowed to control quality, identity or publication.
45
public demo routes
5
named role groups
14
delivery stages
GitHub → Netlify
controlled publication
The problem
Fast output is not the same as credible delivery.
Many local businesses have dated websites, weak mobile journeys and unclear trust signals.
A credible replacement needs research, design judgement, customer-facing copy, suitable imagery, safe enquiry routes, technical checks and controlled deployment.
A general-purpose AI can produce a page quickly. It can also produce a generic, inaccurate or technically valid site that does not feel trustworthy.
The challenge was not to generate more pages. It was to create a repeatable system in which several agents could contribute without losing the audience, approved direction or human standard.
The system
Bounded specialists move the work forward.
I built Site Revamp as a staged operating system rather than a website generator.
Project manager
Pie
Sequences campaigns, briefs and review stages so work moves in the right order.
Lead builder
Claw
Coordinates research, implementation and delivery from the approved direction.
Bounded specialists
Build agents
Implement defined parts of the site without changing the approved brief or release rules.
Independent reviewers
QA agents
Inspect routes, content, forms, mobile behaviour, assets and visual quality.
The agents do not all start from the same vague prompt. They receive the relevant research, brief, selected direction, usable assets, standards, prior feedback and release boundaries.
A critical distinction
A design mockup is not a website asset.
Design mockup
Defines layout, colour, typography, hierarchy and tone. It is a visual instruction for the builder.
Content asset
An image, illustration, logo or other file that may actually appear in the finished site.
James approves the direction and usable assets separately. An early concept cannot become an accidental final decision.
The workflow
Research first. Release last.
Every stage has a defined input, output and reviewer.
- 01
Choose a trade and region with a clear local-business use case.
- 02
Research the intended customer, weak journeys, trust needs and strong patterns.
- 03
Turn the research into a maintained campaign brief.
- 04
Define materially different visual directions.
- 05
Generate design mockups for human review.
- 06
Generate or select content assets separately.
- 07
Let James approve the direction and usable assets.
- 08
Give build agents the approved brief, standards and boundaries.
- 09
Build a static, responsive customer-facing website.
- 10
Run route, metadata, mobile, asset and form-safety checks.
- 11
Submit the completed build to independent visual and audience review.
- 12
Return generic, inaccurate or off-brief work for revision.
- 13
Commit the approved result through GitHub.
- 14
Deploy to Netlify only after the release gate is passed.
The operating model
Four layers keep delivery accountable.
Tools do the work. The operating model decides whether the work can move forward.
People
James decides what the service promises, which direction is credible and whether a result can be published. The system cannot contact prospects, adopt a real identity or make client promises.
Agents
Pie, Claw, build agents and QA agents each have a bounded role, expected output and reviewer. No agent researches, builds, assesses and approves its own work.
Context
Shared Markdown briefs preserve research, audience needs, design decisions, selected mockups, content assets, standards and prior feedback between roles.
Governance
Human design approval, separate asset decisions, automated checks, independent QA, explicit review states and controlled deployment keep public action accountable.
Visible evidence
The output can be inspected.
The gallery contains 45 demo routes across trades, home services, professional services, personal services and health-related local businesses.
Every example is fictional and demo-safe. Names, reviews, accreditations, claims and enquiry details are not presented as evidence of real client work.
Governance in practice
Finished does not mean approved.
Build completeIndependent review required
Builds were returned when they passed technical checks but still looked generic, used weak imagery, spoke to the business owner instead of the business owner's customer, departed from the selected direction or showed a stale preview.
The relevant stage then ran again. The build was not allowed to pass because time had already been spent on it.
That is the point of an independent gate: the agent completing the task cannot decide that its own output is good enough.
Controlled publication
GitHub records the work. Netlify publishes the approved result.
Static builds move through GitHub so the change history remains inspectable. Netlify provides the public deployment layer.
The technical route does not replace the release decision. Real identities, domains, forms, client claims, outreach and publication stay under human control.
Approved build→GitHub→Human release gate→Netlify
What this proves
The strongest proof is the system around the pages.
Site Revamp shows how several AI agents can contribute to one outcome while a person retains control of scope, quality and public action.
Maintained context keeps the research, audience, visual decision, selected assets and previous feedback alive between roles.
Governance provides the definition of good, independent checks and approval gates that stop a fast AI build from becoming the wrong result.
This is the same problem businesses face when they move from isolated AI experiments to repeatable delivery work.
What I learned
Reliable delivery needs more than a capable model.
- One general-purpose agent is not a delivery process.
- Each specialist needs a clear role, input, expected output and reviewer.
- Visual approval must happen before detailed build when design quality matters.
- Design mockups and content assets require separate decisions.
- A customer-facing site must speak to the customer's customer.
- Technical completion is not quality approval.
- Shared feedback prevents every agent from repeating the same mistake.
- Human gates are essential around identity, claims and publication.
Current state
Live proof of concept, not a commercial success claim.
Site Revamp is live as a service site and proof-of-work asset. Its public gallery contains 45 inspectable demos.
The service has not officially launched and has no customer or revenue evidence claimed here. This case study proves the delivery workflow and visible output, not demand or commercial scale.
Next step
Need several AI agents to produce one accountable result?
I can help define the roles, maintained context, quality checks and human gates around one real delivery workflow.