Experiments

My first rewrite of Guildhall passed every test and still looked worse than the version it replaced. The second one worked. The difference was not a better model or a better prompt. It was a Studio: tools that let me, and the AI, see, tweak and compare before anything touched the real world. Part 1 of the Building Guildhall series.

Map generation was the hardest part of Guildhall. It was restarted twice, every frontier model I tried produced the same three mistakes, and the fix was not a better prompt. It was a few engineering ideas: hexagons, relief from stacked layers, sites before roads before nature, density rules, and growth that never moves old land. Part 2 of the Building Guildhall series.

In school I built a Super Mario-style game in Unity for extra credit. That is where I learned colliders, layers, triggers and physics steps. Years later, Guildhall has no Unity and no physics engine, and those same ideas are what kept its townsfolk out of the walls. Part 3 of the Building Guildhall series.

Every animation bug in Guildhall passed its tests. Shields mounted backwards on every clip, legs whirling at 2.3× speed, lamps that lit before anyone touched them, a dragon that flew like a kite. Models wrote motion code that ran on the first try. Making it look right took eyes, timing and a lot of looking. Part 4 of the Building Guildhall series.

Guildhall draws tens of thousands of low-poly pieces, a crowd, a sea and a turning sun in the browser. Getting it smooth started with a list of ideas from experience. Then a measurement contract decided which of them were real: measure on a quiet machine, find the real cost, and keep only the wins you cannot see. Part 5 of the Building Guildhall series.

Guildhall looks like a game, but it ships as a React and TypeScript app. The trick is keeping the two apart: three.js runs its own frame loop outside React, the world is a pure function of a seed, and every rule is a type or a test. That separation is also what let many AI agents work on it at once without breaking it. Part 6, the finale of the Building Guildhall series.

A field map of the LangChain ecosystem after the 1.0 reset — what each piece is for, when to reach for it, and how the parts fit into a real AI product. Part 1 of a hands-on TypeScript series.

Build a real support-triage agent with LangChain 1.0, then layer middleware — summarization, human-in-the-loop, PII redaction, and a custom guardrail — to get production control over the agent loop. Part 2 of a hands-on TypeScript series.

Drop below createAgent into raw LangGraph — StateGraph, cycles, checkpointing, and durable human-in-the-loop resume. We rebuild the support agent as an explicit graph and watch the black box open up. Part 3 (finale) of a hands-on TypeScript series.