DREW
site-agent signalwebgl bayer dither // live context
listening field / no face / no gradient orbfragment shader

Drew Royster · Utah · agents for real work

I build agents that do real work.

Voice interfaces, tool workflows, and local models that turn intent into action.

1. Find the frontier

new capabilities > more autonomy > real work

2. Make it legible

context + tools + traces + bounded actions

Things I do

Agent runtime

The model call is usually the easy part. The harder bit is what it gets to know, what it can touch, and how you catch it when it drifts.

Tool workflows

A model that can't touch the system is mostly a text box with confidence. The useful version calls tools, stays bounded, and leaves enough evidence to check.

Voice interfaces

Voice is unforgiving. A little lag makes the whole thing feel dead. Bad interruption handling turns a smart model into a weird phone tree.

Local systems

Some things belong close to the machine: private data, local models, home servers, weird hardware. Cloud is useful. It just isn't always the default.

Automations

Good automations should be boring. They wake up, check the world, do the thing, and leave a trace you can inspect later.

Evaluation loops

If you can't inspect it, you're mostly judging vibes. Tests help, but so do transcripts, screenshots, diffs, and screens that make failure obvious.

Work records

Systems that survived contact with real work. Each record keeps the constraint that made it interesting.

tools and machines

Staff Search

Search that had to survive messy people data

The interesting part was not search in the abstract. It was getting messy people data clean enough that an operator could ask a normal question and get useful candidates back.

Constraint: translate fuzzy operator intent into inspectable retrieval and function-calling steps over messy structured and unstructured records.

Open record

Shaolin AI

Learning the fundamentals by training models

A bootcamp I attended as both a student and substitute teacher to upskill, where I trained models, got much more comfortable with PyTorch, and built a movie recommender algorithm for my capstone.

Constraint: get past AI product demos and understand the machinery directly enough to train, debug, and explain small models.

Open record

Smart Autofill

Forms are still where the pain hides

A browser extension experiment for using personal context to fill out forms. It sounds small, but forms are where a surprising amount of real workflow pain piles up.

Constraint: make personal context useful inside a browser workflow without letting the model invent facts or overreach.

Open record

Field Notes

Note 02 · November 17, 2025

Legibility

You know things (allegedly) how can you make your agent know them too.

Read article