Building This Site with AI Agents: A Meta Journey
You're reading this on the Little Research Lab. And the Little Research Lab was built by AI agents. Which makes this post a bit like a snake eating its own tail, but with better test coverage.
The Setup
I wanted a simple publishing platform. Somewhere to share research, host PDFs, track who's actually reading what - the usual. But I also wanted to see how far I could push AI-assisted development using a squad of specialized agents instead of one generalist doing everything.
The cast of characters:
- Solution Designer - scoped the project before anyone got too excited
- Business Analyst - produced specs, tasklists, and quality gates with the enthusiasm of someone who genuinely loves acceptance criteria
- Coding Agent - wrote the actual code (the star, arguably)
- QA Reviewer - the one who noticed when things were quietly falling apart
The Process (and the Chaos)
The workflow was deceptively simple: Solution Designer sketches the boundaries, BA produces detailed specs with test assertions, Coding Agent implements, QA reviews. Repeat until done.
What actually happened: the agents produced a 377-line spec document before anyone wrote a single function. There's a section called "Unknown-unknowns checklist" with risk controls. The spec includes DST-safe timezone handling for scheduling. For a blog.
The BA agent created quality gates that would block deployment if any draft content could accidentally leak to the public. The spec mandates "zero audience PII" - no cookies, no tracking pixels, no IP logging. These weren't my requirements. The agents just... added them. Because apparently they've read about GDPR.
When Things Got Interesting
Halfway through, the QA agent flagged a problem: the codebase was 36% compliant with the spec's "atomic component" architecture. There were 61 type errors and 58 linting failures. The shell layer was importing from deprecated service locations.
So the agents ran a four-phase remediation:
1. Deprecate the old code (with a polite DEPRECATED.md explaining why)
2. Migrate to the new pattern
3. Fix imports across the codebase
4. Clean up the debris
The evolution log shows this as "EV-0002: AC-TA mismatch / incomplete migration / split-brain architecture." Resolved. Evidence: mypy 0 errors, ruff 0 errors.
The agents kept receipts.
What I Learned
AI agents are excellent at generating bureaucracy. I mean this as a compliment. The spec files, decision logs, and quality gates created a paper trail that made debugging straightforward and decisions reversible.
But the real surprise was how opinionated the agents became. They added security features I hadn't requested, structured the code for maintainability I might never need, and documented everything like they were preparing for a code audit.
Building software with AI agents feels less like pair programming and more like managing a small, extremely thorough consultancy that never sleeps, never complains, and occasionally rewrites your architecture because it found a better pattern.
The repo is public now. Judge for yourself.