Agentic AI Experiments

Overview

Most of what’s on this site - the projects, the write-ups, the site itself - shipped through Claude Code agent teams. This page is the running log: how the setup works, what I’ve adopted, what got thrown out, and where it stands.

Why agents

It started as curiosity about whether the tools were actually good enough to trust with real work, and it’s turned into the biggest change in how I build things since I learned git.

The team structure

The top-level session acts as a principal engineer. It owns the architecture, makes the cross-cutting calls, and delegates bounded chunks of work to senior-engineer subagents running on a lighter model. A senior owns its piece end to end, makes its own implementation decisions, and reports back the judgment calls it made along the way. That last part is the load-bearing detail: reviewing judgment calls is much cheaper than reviewing diffs, and it’s usually where the bugs hide.

I tried a three-tier version with a junior-engineer role for mechanical single-file edits. It got retired. Decomposing work small enough for a junior cost more than the junior saved, so two tiers is where it settled.

A few conventions hold the whole thing together:

• Project repos carry a START_HERE.md that any agent reads before touching code, so conventions get followed instead of guessed at.
• A persistent memory carries feedback across sessions - standing rules like “no em dashes” and “verify UI on mobile and desktop before calling it done.” I complain once and it sticks.
• This repo has a VOICE.md that defines how anything shipping on the site should read. Agents draft, I red-pen.

The toolbox

Plugins I’ve adopted, roughly ordered by how much they’ve earned their keep:

• superpowers - process skills more than tools: brainstorming before building, test-driven development, systematic debugging, and a verification gate that demands evidence before an agent gets to claim something works.
• code-review - multi-agent review of a branch before it merges.
• feature-dev / code-simplifier / claude-md-management / hookify - guided feature work, post-hoc simplification passes, keeping project instructions honest, and turning “stop doing that” into an enforced hook instead of a hope.
• frontend-design - the newest install, aimed at the visual work on this site.

A few more are installed and unproven. I’d rather list the ones that changed how I work than inventory a plugins folder.

At work

At Medtronic the sanctioned tool is Copilot CLI, so that’s what I use there - and I do my best to treat it exactly the way I treat Claude Code at home: scoped delegation, explicit specs, review before trust. The discipline transfers even when the tool changes. That’s what convinced me this is a way of working, not a feature of one product.

Status

This is a running log, not a finished build - it stays open. I direct and make the calls; agents implement, review, and record. I trust the workflow enough that it built the site you’re reading, and I distrust it enough to keep every gate in place.

Results

Workflows that stuck:

• Staged gates. The Robotic DnD demo was built this way: camera and projection, perception, grasping, video capture - each gate with its own spec, implementation plan, and review before the next one opened. No gate opens on top of unreviewed work.
• Cross-review. Agents review each other’s output before I see it. My review then happens at the level of intent and judgment, not syntax.
• A knowledge base of transferable findings. Hard-won lessons - USD physics schemas that silently no-op, cameras that render black without a light and a warm-up frame - get recorded where the next agent will find them, instead of being relearned per session.
• Negative results get written down with the same care as wins. That’s a rule, not a preference, and it’s the reason the write-ups on this site admit what doesn’t work.
• Parallel teams. Independent chunks run simultaneously. The contributions carousel and the project preview modals on this site were built by two agent teams in the same afternoon while I made the design calls between them.

Challenges

The failure modes are real - the carousel on the homepage shipped working on desktop and broken on mobile because verification only covered one screen size. The fix came with a new standing rule, which is the pattern this whole page keeps repeating: the system gets better by writing down what went wrong.

What’s next

A few plugins are still installed and unproven - whether they earn a place in the list above is still open.

This page included. An agent drafted it, I edited it. That’s the experiment.