AI agent pipeline
Website-Generator
An 11-stage agent pipeline that turns business data into a checked, static website.
In development, first complete run on 7 Aug 2026
Research
Intake- Social
- Video
- Press
- Competitors
Planning
Planner- Call from the voice agent
Build
Builder- Reviewers
Gates
Privacy text- Third-party content
- Accessibility
- AI labelling
- Image rights
Publishing
Go live
Agents build a website
Watch a business's data turn into a checked website, step by step.
A simulated run with a made-up business. The stages and gates are real, their order is simplified, and the gate that fails here is an example.
Core skills
- LLMs and agents
- Security and privacy
- Reliability
- Infrastructure and ops
- Data engineering
- Frontend
- AI-assisted engineering
Tech stack
- Claude Code (headless)
- Bash
- Cron
- Next.js 16
- React 19
- Tailwind 4
- nginx
- Playwright
- sharp
- ffmpeg
- whisper
The hardest part
The website with placeholders in its legal notice
An order that only exists in the plan isn't a rule.
The whole story
What it looked like
On the first complete run, the finished site had seven placeholders in its legal notice. The answers were only a three-minute phone call away.
What was really going on
The planner and the call were skipped, and the builder went straight ahead. The eight steps were started by hand in any order, because the right one was only in the plan, not in the code.
The fix
An orchestrator now enforces the order: no call, no build, and if a required step fails, the run stops. It then resumes from that point instead of spending another 23 minutes transcribing a long podcast.
- 11
- pipeline stages
- 57
- commits in the generated sites repo
- 13
- classes of artefact in a site teardown
In plain words
A small business needs a website but has no time for one. Several AI agents gather information about the business, plan the site, build it and check each other's work. It only goes live if automatic checks agree. After that, you request changes by phone.
A website is expensive for a small business and goes stale quickly. The pipeline automates the whole process, and the agents don't just produce text but have to pass hard checks.
How it works
Step 1 of 5
Gather
Intake, social, video, press and competitor stages gather what is known about the business.
What was built
- 11 stages: intake, social, video, press, competitors, planning, call from the voice agent, build, reviews, gates, publish
- Gates for privacy text, third-party content before click, accessibility, AI labelling and image rights
- Static Next.js output behind a generated nginx vhost, with an on-server rehearsal and rollback
- Change loop: the customer calls, agents implement, a callback confirms, then it goes live
- A contact form receiver that stores, mails, retries and deletes after 90 days
- Site teardown in a fixed order across 13 classes of artefact, with step-up auth and an audit trail
How those skills were used
- LLMs and agents
- Multi-stage orchestration of headless Claude Code runs with review stages and gates
- Security and privacy
- Sealed instruction files, so one customer's agent can't plant instructions for another's
- Reliability
- Cron workers with locks, stale-run recovery and clear terminal states
- Infrastructure and ops
- The nginx config is generated at publish time, rehearsed on the server and can be rolled back
- Data engineering
- Scraping and media pipelines with Playwright, sharp, ffmpeg and whisper
- Frontend
- Generated Next.js sites and a multi-language marketing site with three.js elements
- AI-assisted engineering
- Seven review rounds on the instruction fix, in which the first approach was thrown out entirely
My role
Concept, architecture, directing the agents and verification. Claude Code wrote most of the code, inside an agent workflow I developed myself.
See the Context Observer project