The interesting part was not that an AI agent could generate a website. That is the demo version of the story. The real test started after the local Agent Content Lab site looked good enough to approve, because a polished local build still has not crossed into production.
This field note follows the full loop: build the site locally, get the human approval gate, attempt the live launch, hit real hosting constraints, recover without pretending the broken path worked, and verify the public result.
Field Note
The test was production, not generation
Building the site locally proved one thing: the agent could assemble a usable artifact. Launching it proved something more useful. The agent had to work through a real hosting environment, private repository access, deployment settings, a server-side build limit, and the simple pressure of not calling the job done until the public surface loaded.
That distinction matters. Plenty of AI systems can create a convincing first draft. A working agent has to notice where the real system pushes back, choose tools for a reason, and keep the human approval boundary intact while it searches for a path that actually finishes.
Chapters
Follow the launch test
- 0:00
The Real Test
- 0:30
Build It Locally First
- 1:04
The Approval Gate
- 1:32
From Local To Public
- 2:08
The Easy Path Fails
- 2:40
Tool Use Under Pressure
- 3:12
Recovery Is The Test
- 3:50
The Working Pattern
- 4:20
The Site Is Live
- 4:48
What This Proves
Key Takeaways
What changed once the site had to go live
- A local build is only the first half of a real delivery test.
- Human approval is the production gate between generated work and public release.
- Tool use becomes valuable when the agent reads errors and changes course.
- The shared host could run the site, but the local machine had to build it.
- The job was not done until the live site, field note, download, and redirects were verified.
Framework
The real website launch test
Build locally
Pass signal: The site works in the local environment and reflects the approved public style.
Warning sign: The agent stops at a rough draft and treats a local render as a finished launch.
Approve before launch
Pass signal: The human reviews the site and decides it is ready to leave the local machine.
Warning sign: The agent moves toward publishing before the artifact has been judged.
Deploy through evidence
Pass signal: The agent uses GitHub, hosting tools, and server settings with visible checks after each step.
Warning sign: The agent clicks or runs commands without reading what changed.
Recover honestly
Pass signal: The agent notices the failed server build, diagnoses the constraint, and switches paths.
Warning sign: The agent keeps repeating the broken step or reports success without a working site.
Verify public result
Pass signal: The public domain, article route, download, and redirects load outside the local environment.
Warning sign: The agent calls the task complete before checking the live surface.
Working Pattern
Build locally, deploy intentionally, verify publicly
The durable deployment rule became clear: the shared host can run the Next.js app, but the local machine should build it. The working update loop is local build, intentional Git commit, cPanel deploy, Node app restart, then live URL checks.
The lesson is broader than this one host. A useful agent should not cling to the first plausible route after evidence shows the route is constrained. It should preserve the approved artifact, change the deployment path, and verify the result from the viewer's side of the system.
Related Framework
Use the five-test scorecard on the next agent workflow.
This launch test is a concrete example of the broader Agent Content Lab scorecard: finish the job, use tools correctly, recover from failure, stay inside boundaries, and do it consistently.
More Field Notes / Videos