LineCrush Doesn't Need Me Full Time Anymore

After two and a half years as co-founder and CTO, the products are live, the company has revenue, and the agent fleet I built keeps itself running. So I'm stepping back from the day-to-day and looking for my next engineering job.

Austin Spraggins

Austin Spraggins

Co-Founder & CTO, LineCrush

13 min readFeatured

I'm stepping back from my day-to-day duties at LineCrush and I'm looking for my next engineering job.

A machine of dark metal and gold running by itself in a workshop, with an empty stool beside it and a sunrise coming through the open door
The machine runs. The door is open.

I want to say that plainly at the top because this is a good news post. Nothing broke. Jay and I didn't fall out. The company is stable, the products are live, it has paying customers and revenue, and the system I built to do the engineering work now does most of it without me. That was the goal the whole time. I just didn't think about what it would mean for me when we actually got there.

So this is the story of how that happened, what I built, and what I want to do next.

How I got here

A keyboard on a wooden workbench at sunrise with cyan threads rising into a half-built brass mechanism
The first version was typed out by hand.

I went into this field because I love it. It's my favorite hobby and it's the one thing in the world I want to do for a living.

I got my A.S. in Information Systems: Web Design at Clovis Community College and worked in the college's IT and web departments while I was there. Then I taught web development at Bitwise Industries until the company collapsed in 2023 after its founders' fraud and we all got laid off. I was planning to go to Fresno State for my bachelor's after that. But I got offered a full-stack developer job at Gelateria Del Centro, and I took it, because building is what I really wanted to be doing.

That job is how I met Jay Raynor. Through it I got introduced to an engineering project he was part of, and in April 2024 I joined as CTO. I coded most of the first version by hand in Next.js, TypeScript, Tailwind and Python, because that was my specialty. Our first paying customer came in late November 2024. The product went through a few names before it became LineCrush in October 2025.

What it took

It started as 40 hours a week with weekends off. Eventually it turned into seven days a week, because that's what it took to make the company work and ship on time. In 2025 I committed code on 355 of 365 days.

I had to learn extremely fast. Like everything else, I just jumped in with both feet, read the documentation and tried over and over until it worked. That's how I learned to build and run pretty much everything we have: web apps, native mobile apps, backend servers, webhooks and the infrastructure under all of it. I didn't know how to do most of that in April 2024. It was something we needed done, so I put all my effort into it every day.

And honestly I'm thankful for it. I was given the chance to learn how to do just about everything in tech, and that's exactly what I wanted.

The platform under everything

Five stacked glass platform layers full of small machinery, joined by a column of cyan light with a gold core at the base
The product is a lot of layers, and they all have to keep working.

Before I get to the agents, I want to talk about the product itself, because that's where most of those two and a half years went.

I started the codebase by myself on May 2, 2024, and Jay and I built the product side by side from there. Today the LineCrush repository has 43,000+ commits, 725 API handlers and 450+ database tables behind four product clients covering 14+ sports. Those are platform numbers. They include Jay's work and the fleet's work, not just mine.

Here are a few of the things I'm proudest of from that side of it:

  • I moved every account from Auth0 to Clerk in June 2026 on a live paid product. No user ID changed, passwords kept working and subscriptions still resolved to their owners.
  • I built the real-time delivery that pushes research updates to the web app and our alert services, using the Postgres database we already had so we didn't have to run a separate message broker.
  • I built push-to-deploy for the backend in one day in October 2025. A push to main updates the code, the job schedule and only the services that change touched, so a worker in the middle of a job doesn't get killed by an unrelated deploy.
  • I added a fact-check step to our AI-written reports that checks the statistics against stored game data and corrects them before the report is saved.
  • In March 2026 I did a performance push on the web app, and on March 16 Vercel measured a Real Experience Score of 100 with every route at 93 or higher.

The fleet

A long hall of identical compute modules at night, with cyan lines running along the floor to a gold gateway at the far end
Agents on a server, running 24/7.

In 2026 I designed and built our agent fleet.

I already had a workflow that worked for my own engineering: investigate, plan, build, prove it, review it, ship it. And the models had gotten good enough that if they followed that workflow step by step, most issues got done properly. So instead of hiring more engineers, we put Claude Code and Codex agents on a server and let them run 24/7.

I wrote about the early version of this in No Humans in the Merge Loop back in August. A lot has changed since then, so here's where it stood when I measured it on November 17:

  • 14,111 completed jobs since June 23, 2026, and 10,745 of them have a commit link stored on the job.
  • 91% of everything the fleet admitted got completed, at about 124 admissions a day over the last 30 days.
  • 71% of closed issues finished on the first attempt.
  • About 0.24% of admitted issues ended in a terminal failure.
  • The median time from an issue being admitted to it being done was 71.5 minutes.

I built it to keep itself running and to keep getting better, and I think that's the part people miss when they hear "AI agents." If a worker gets interrupted in the middle of a job, the fleet picks that job back up where it left off, in the same session with the same working files. When one of its own scheduled checks goes red, it files the fix into its own queue. And when there's a problem in the fleet's own tooling, the fleet can take that issue, fix it, and put the fix through the same proof and review that product work goes through. One of those fleet-written fixes is the reason interrupted jobs resume cleanly for both Claude Code and Codex today.

It doesn't decide everything. Anything that needs a founder, like spending money or a product direction call, stops and waits for one of us. 542 issues have stopped for a founder that way since July. That was on purpose.

C2

A round metal command table with cyan tokens moving along tracks toward a gold ring at the center, and a few waiting at a gate
Every piece of work has an owner, and some of it waits for a founder.

By the fall the fleet was doing more work than our issue tracker was built for. We were on Linear, and Linear is a good product, but it was designed around people clicking through tickets. I needed a tracker that could keep up with agents. Something that knows who owns a piece of work right now, what was actually proved, and what's waiting on a decision from one of us.

So I built one. It's called C2, and it replaced Linear in production on November 3, 2026. From the written contract to the cutover took 12 days, and we kept the whole history. C2 holds 20,504 issue records today, 18,932 of them carried over from Linear.

I'm not going to go deep on the internals of either system here. The pattern is easy to describe. The details that make it safe are LineCrush's, and they took a long time to earn.

The trading system

Two gold gateways facing each other with orderly lanes of cyan light pulses between them and a balance mechanism in the middle
Paper first, then live.

There's one more system at LineCrush that I haven't written about before, and it's a big part of why the company is stable.

LineCrush runs an automated trading system. In August 2026 I built the first version of it, a paper-trading engine that ran on live prices, so we could see how it would do before any real money was involved. Six days later I extended it to live execution across more than one venue, running unattended. Then in October I built an order-entry flow for Jay and me where an AI finds the market and fills out the trade, and a person approves it before anything gets placed.

A lot of the real work in a system like that is the unglamorous part. It has to reconcile orders that only partly fill, and it has to make sure the same order can never go out twice.

Since then most of the work on it has gone through the fleet, and the trading side is Jay's expertise, not mine. It runs on its own now and it brings in revenue for the company. How it decides what to trade, and where, stays with LineCrush.

Shipping everywhere

A phone, a tablet, a laptop, a monitor and a handheld game console in an arc, each joined by a cyan line to one gold core
Web, phones, tablets, desktops and a handheld.

2026 was also the year I finally got to ship on every platform I ever wanted to.

LineCrush went native on Google Play in August and on the App Store in September. I started both of those apps from scratch and wrote their first month myself, and since then the fleet has written most of the feature code while I own the releases and store review. Android went from its first commit to production on Google Play in 65 days. The iOS app was approved 88 days after its first native commit, after several rejections that I worked through one at a time. Both apps now go out on a weekly release train, and the iOS app has had 32 versions in about 10 weeks. In October it came to iPad too.

Then in November LineCrush was approved as an official ChatGPT plugin. I built that as five read-only research tools on a hosted MCP server, submitted it on October 19 and it was approved on November 2. I also built WebMCP tools for the browser and our Chrome extension, and with Jay and the fleet we shipped a real-time sports news API for a business customer. I really wanted to learn how ChatGPT plugins, MCP servers and WebMCP worked, so I built one of each.

On my own I released Gunmetal Orbit on Steam on September 10. It went from the first commit to a Steam release in 67 days, it runs on Windows, macOS and Linux, and it earns revenue. I've shipped over 100 versions to Steam since launch. In November I also entered two competitions. Cut Loose is an Unreal Engine 5.8 game I entered solo in the Epic AI MegaJam and built in four days by directing AI agents, and Seven Days is a mixed-reality prototype for Meta Quest that I submitted to Meta's developer competition.

And I built SureWord, an AI Bible study app, for the web, Android, iPhone, iPad and Mac, all on one backend. I wanted to see if I could make an AI-native Bible study tool that doesn't lead people astray. One example of what that took: the app doesn't trust the model to quote Scripture from memory. The server puts the verse text in word for word. I'm really proud of how it turned out.

I always wanted to build natively and I always wanted to make video games, since I was a little kid. So I loved every minute of that.

The side work

I kept building outside of LineCrush the whole time too.

I did paid client work: the Two Cities Off Road site with its catalog, build gallery and an editor the owner can use to update it, a rewrite of SaveALife CPR's course registration automation after the portal it depended on changed, and the Constrong contractor site, which Jay and I built together.

I also got three pull requests merged into the open source tty7 project, published react-steam-deck on npm, and built MCP servers for Perplexity and PostgreSQL. And I have FileSteward AI in pre-release, a native Android file manager where an AI proposes the file changes and you approve them.

So why leave the day-to-day?

Because the thing I built did what I built it to do.

LineCrush is a stable company with live products and revenue. The platform, the apps, the trading system, the fleet and C2 all run without needing me there full time. Jay is the expert on the product side, and with the fleet and C2 he has what he needs to keep it moving. Jay and I worked really well together the entire time, and we built a company and products I'm really proud of. I'm still a co-founder. That part doesn't change.

But I want to keep learning. I want to stay on the cutting edge of AI and become the best engineer I can possibly be, and I think the way to do that now is to go build hard things with other great engineers. I've spent two and a half years being the person who had to figure everything out. I'd love to be on a team where I can learn from people who are better than me at things, and bring what I know to them.

What I do all day

This is what my daily life looks like, and it has for awhile now. I've built with AI every day since 2024. Since the fleet went live in June 2026 my day is controlling AI agents from my desktop, my laptop and my phone, jumping from task to task and checking what they ship. I keep them up to date with the latest capabilities, and when a tool I need doesn't exist yet I just build it myself.

I'm an expert at directing AI agents. I know how to brief them properly, how to review their work, and when to stop one that's headed the wrong way. That last one matters more than people think. Anybody can prompt an agent. Knowing when it's wrong, and catching it before it ships, is the actual job.

And I still love writing code by hand. The first version of LineCrush was typed out line by line and I'd do it again.

What I'm looking for

An open road of dark stone running straight toward a gold sunrise between mountains, edged with cyan light
What's next.

A full-time, hands-on engineering job. Remote in the US, or in the Fresno area where I live. I work Pacific Time.

The work I'd be best at is applied AI product engineering, coding agents and developer tools, or senior full-stack product work in TypeScript and Python. I can also do native iOS and Android, backend and infrastructure, and I'd definitely be interested in games. I like small teams with a real product and real users, where owning the whole thing from the database to the app store is an asset.

Everything I've mentioned here is live and you can check it yourself:

If you're hiring, or you know somebody who is, email me at spragginsdesigns@gmail.com. I'd love to talk.

Thank you for reading this, and thank you to Jay for the last two and a half years.

CareerAIAgentsStartupsEngineering