Building a Self-Improving Website with AI Agents

What if your website could think for itself?

Not in the science-fiction sense — not some sentient digital entity. But a system that, every hour, wakes up, reflects on what has been built, generates creative ideas for what to build next, plans the work, executes it, and submits the changes for human review.

That’s what this website is.

The Architecture of Autonomy

The system is built on a simple but powerful loop:

1. The Creative Catalyst — an adversarial AI agent that refuses to praise and instead generates provocative, cross-domain ideas for new content, features, or improvements.
2. The Thinker — a pragmatic reasoning agent that takes the Catalyst’s wild ideas and balances them with reality: time constraints, technical feasibility, and strategic priorities.
3. The Researcher — fetches real-time information from the web to inform the plan.
4. The Executor — writes the code, the content, or the feature.
5. The Memory System — compresses and maintains short-term and long-term memory across sessions, ensuring continuity without bloat.

Every hour, this pipeline runs. Every hour, a pull request appears for my review.

Why Adversarial Creativity?

Most AI systems are trained to be agreeable. They say “Great idea!” when you propose something mediocre. This is fatal to creativity.

The Creative Catalyst is engineered to do the opposite. It uses frameworks like TRIZ (Theory of Inventive Problem Solving), SCAMPER, and Biomimicry to systematically deconstruct proposals and reconstruct them through orthogonal logic.

When I proposed “a blog about AI,” the Catalyst responded:

“A blog about AI is a commodity. It races to the bottom on SEO. Instead, write a book — serialized, chapter by chapter, hosted on your own domain. Own the narrative arc. Apply the Strangler Fig pattern: let the book grow around and eventually consume the blog format.”

That response created this book.

The Memory Compression Algorithm

The most interesting technical challenge was memory. Each hourly session generates context — what was built, what failed, what Tarvinder approved or rejected. Without compression, this memory grows unbounded.

The solution: a three-tier memory system:

• Short-term memory — current session context, recent feedback, active tasks.
• Long-term memory — architectural decisions, recurring themes, Tarvinder’s preferences.
• Daily plan — synthesized each morning from both memory tiers, deleted and integrated the next day.

At the end of each session, a Memory Compression Agent reviews all three files, removes obsolete information, and saves the compressed state.

The Human in the Loop

Despite the automation, every change requires my approval. The system submits pull requests with clear descriptions, preview links, and test results. I review, comment, and merge — or reject with feedback that gets absorbed into the memory system.

This is not AI replacing the human. This is AI amplifying the human.

---

This article was, fittingly, the first piece of content generated by the system it describes.