Jackson Cantrell’s Weblog

A few people, a lot of agents, and who makes the calls.

Five levels of self-driving, for code

Pixel art of a teal convertible climbing a switchback road in the Verdugo Hills at sunset. The driver has both hands on the steering wheel. The road zigzags up a chaparral hillside below a ridge topped with radio towers, the sun sets behind faint purple mountains, and city lights are coming on in the valley.
Level five, the way I’d build it: the car does the climbing, and somebody keeps a hand on the wheel.

In February I went back on Barron’s Streetwise, the show I used to produce, to talk with Jack Hough about AI coding. To explain what had happened to me, I borrowed the levels of self-driving from cars and applied them to writing software. This is that framework, written down, and where I’d put us now.

First, where I started. About a year before that episode, I was hopeless at coding. I’d taken one programming class in college. I spent something like 120 hours on the final project, a Flappy Bird clone that didn’t work very well, and I barely passed.

Level 1: autocomplete

The first AI coding tools finished the line you were typing. For a working programmer, that’s a real time-saver. For me it was useless, because I didn’t know what the suggestions meant. You can’t accept or reject a suggestion you can’t read.

Level 2: describe it, get a front end

The first tool that changed things for me was Vercel’s v0, which turns a description into front-end code. At Dyrt we needed a dashboard for customers to see their food waste. The usual path was weeks of back-and-forth with a contract developer over what it should look like. With v0, as I put it on the show, “I was able to just whisper into this machine and it would come out with a result that was much closer to what I wanted.” If I wanted a tweak, it happened right away. Then the developer connected it to real data in about an hour.

That was the first time I could make the thing in my head show up on a screen without going through someone else.

Level 3: a chatbot in the editor

Next was Cursor: a chatbot inside a code editor. At this level you could build apps that actually worked, but you needed to know something about code to keep it honest. It never pushed back. It would never tell you that you were being an idiot, even when you were, and it happily built inefficient ideas exactly as described. It also wrote new code every time you asked, so you could end up with the same function written five different ways in five different places.

Level 3 rewards people who already know what good code looks like.

Level 4: describe the product, get the product

Around Christmas 2025 I started using Claude Code with Opus 4.5, and the ground moved. You could know very little about code, describe what you wanted, and conjure a working piece of software into existence, as long as the scope was small. I said on the show that you’d never have to look at the code. For small things, that was close to true.

Level 5: a town of agents

Level 5 is going from one agent to a team of them. For me that was Gas Town, Steve Yegge’s framework, which is themed after Mad Max. There’s a mayor who coordinates. There are polecats, worker agents that each build one feature. There’s a dog that wakes up polecats that have gone off track, and a witness that writes everything down. You give the town a plan and it works all night.

In February I said I couldn’t recommend it to most people, because running that many agents was expensive. And on the whole scale, I said we were at about a 4.1.

A pixel-art staircase of the five levels of AI coding, each step with a small scene on top. 1, Autocomplete, finishes your line: an old beige monitor. 2, v0, describe it, get a front end: a screen with a dashboard of charts. 3, Cursor, a chatbot in the editor: a laptop with code beside chat bubbles. 4, Claude Code, describe it, get an app: a terminal with a finished app window. 5, A town of agents, keeps the wheel: a small town at night with a town hall, lit houses and little workers, and a person at a steering wheel at its edge. A flag beside step 4 marks 4.1 in February 2026, with a higher pennant reading 4.? for now.
The five levels from the show. We were at 4.1 in February; I still won’t guess the decimal.

Where I’d put it now

Higher than 4.1. I won’t pretend to know the decimal.

A few things were already true in February and have only grown since. Claude Code had added agent teams two weeks before we recorded, and working with several agents at once has since stopped being an unusual setup. This site, which Jack poked at on the show, was already rebuilt with Gas Town, a reasonable test of whether someone who barely passed one coding class can ship software now. And my brother Clay and I now run our own agent mail across our machines.

But I’ve also changed my mind about what level 5 should mean. In a car, level 5 means no steering wheel. For software, I don’t want that. The town working all night is the easy part now. The hard part is the morning: someone has to read what it made, decide what’s right, and turn it into a decision a person can make. The more the agents can do, the more that part matters.

So my version of the top level keeps the wheel. The agents do the driving, and a few people keep the decisions. That’s the version I want to help teams build. There’s more about me on the about page, and Jack wrote about the conversation in his column, Vibe Coding in Gas Town.

Jackson Cantrell

Jackson Cantrell is an AI-native CTO and agentic engineering leader in Los Angeles. He helps small teams build software with many AI agents at once while people stay in charge of the decisions, and he helps teams that are new to agentic tools get up to speed.

He and his brother Clay Cantrell were early contributors to Gas Town, which shipped multi-agent orchestration and agent-to-agent mail in January 2026, weeks before Claude Code added agent teams. They run their own agent mail system across machines. Jackson built emBEADings, an open-source tool for the Beads issue tracker that Gas Town runs on.

Before that he was on the founding team at Dyrt, an organic waste startup, where he was hired to lead marketing and went on to lead product. Customers asked for tools, so the team built waste bill analysis that showed true diversion rates and hidden fees, invoice ingestion, waste stream tracking, an integration with Schneider Electric Resource Advisor, and Dyrty Vision, a computer vision system that tells food scraps from wasted food and reports back to food and beverage teams. Dyrt grew from 1,000 to 1.5 million pounds of food waste a month, signed national deals with a grocery chain and a hotel chain, and ran a facility in Vernon, California that employed 30 people.

Dyrt was an attempt to fix a broken California recycling industry. Jackson, Clay and the Dyrt team published an audit of Los Angeles's RecycLA franchise, which found that at most 18 to 24 of every 100 tons of commercial waste could be counted as diverted from landfill. He learned the hard way how well-meaning but burdensome rules and entrenched interests can sink a facility that works: the permit for Dyrt's indoor composting facility in Vernon went before the Vernon City Council in August 2026 and was carried into September. The company ran out of money that fall, and his last day was October 1, 2026.

Earlier he reported on tech and finance as a producer and co-host of Barron's Streetwise with Jack Hough, led audio product at Finimize, and reported for public radio: KCRW in Los Angeles, and North State Public Radio, where he was part of the team whose 2018 Camp Fire coverage won an Edward R. Murrow Award. He studied neuroscience at Brown University.

He is open to conversations, collaborations and new roles: jaxtrell@gmail.com.

How he works

Projects

Work history

Press

Jack Hough, host of Barron's Streetwise, interviewed Jackson about going from no coding to shipping software with AI agents, and about rebuilding this site with Gas Town as a fictional 2002 college radio station.

Stories

Links