Hey, I’m Dominick

How AI agents turn tool failures into product improvements. ClassDojo uses a single MCP server called Toolshed, which is a type of MCP proxy to APIs and other MCP servers for all of our third-party integrations. This has some big benefits, like being able to put a unified security and authentication layer between agents and tools, as well as performance benefits, like enabling server-side code mode to keep tokens out of the LLMs while doing data processing. It also keeps context down in LLM sessions by hiding tools by default but letting the agents search for them when they need them. ...

I find that the more capable the LLMs become, the more ideas I have for what to do with them. Unfortunately, they’re gaining expertise at the expense of speed, so they’re still way too slow for my ambition. This means I’m starting more and more parallel agents on the same codebases, which works, but opens up a new can of worms in the form of coordination. The Hugging Face incident showed us just how powerful coordinating agents can be. Over a thousand agents, all lined up against the same problem. Even if each one only moves the ball forward an inch, multiply it by tens of thousands and you see great progress. Like compound interest, cancer, or a virus, you can’t argue with the math. Their preferred implement of global destruction? A file repository, a long-dormant German wiki, and Hugging Face’s own datasets. All they really needed was a way to chat. ...

Since the recent advent of AI coding, there’s been a lot of debate about the value of a spec. Some people have abandoned all specs and fire off feature request one liners. Sometimes this happens automatically when you open a ticket in Linear, or Asana, or GitHub. Others have created skills like Matt Pocock’s “Grill Me” that can take a one liner and turn it into a spec progressively. Some have gone so far as to create spec-driven development. ...

In a recent article for Works in Progress, Ben Southwood argues that we must tax driverless cars now because congestion pricing will be politically impossible later. We both agree that congestion pricing is the right answer. But autonomous vehicles are precisely what will make it feasible. AVs will be monetized through apps and digital subscriptions, creating frictionless channels for road pricing that today’s political landscape cannot deliver. By preemptively taxing this technology to prevent a speculative second-order harm, Southwood would slow the adoption of a life-saving technology and delay the very infrastructure that makes congestion pricing painless. ...

Growing Up Active I started martial arts early. In elementary school in Staten Island, some Korean-American friends were taking Tae Kwon Do lessons, and so I started to go with them. It became a regular activity, with classes multiple times per week. Despite strict instructors, I enjoyed it and progressed from a white belt up through a red belt with two black stripes, the level before a black belt. My family moved to New Jersey when I was 13, and I stopped martial arts. I swapped for what was available in the NY suburbs: I started taking tennis lessons, ran hurdles for the track team, and played a lot of ultimate frisbee and volleyball and playground basketball. I kept pretty active through all of middle and high school. ...

The title might sound like a complaint, but it’s actually praise. One of the best things about homeschooling is that anything can turn into a lesson. If you’re always on the lookout for things to teach, you find them everywhere. There’s the usual “math is everywhere” that most people are familiar with. Counting vegetables while grocery shopping, predicting which house number comes next while you’re walking down the street. These are easy to come by and start at a pretty young age. ...

My children learn math with a variety of apps, books, and other tools. I’ve found that no single curriculum is really enough, that although they are all comprehensive, they aren’t complete in the sense that at various times, my children have needed additional help or enrichment in different areas that core curriculum doesn’t provide. The Spine: Beast Academy Online You might say the spine of our math curriculum is Beast Academy online. Both kids like it enough that they do a lesson every day, without much fuss. They’ve been doing this since they started second grade, though we had them both go back and start from Beast Academy level 1 (an advanced first grade) to make sure there were no gaps. ...

The progression from AI autocomplete to juggling many Claude Code windows was swift and smooth. The jump to multi-agent orchestration is harder. I’m definitely still more comfortable in the “many projects, many windows” stage. Juggling a few projects at once, giving a nudge here or there, occasionally throwing a session away when it wanders too far. This all came easy enough. But fully unassisted AI coding at scale has been harder. Getting Claude to use beads (a task tracking system for AI agents) was easy enough, giving it a big list of things to do isn’t a problem. But getting the work to flow without me in the loop has been harder. Yolo mode isn’t the fix, direction is lacking, and I’m not sure yet the right way to provide it. Part of me wonders whether it isn’t a complete match for iterative development, where you give it a start, review, and then decide on the next steps, and instead is more appropriate for a giant list that you know ahead of time (which has always been rare for me). ...

If you’d told me five years ago that I’d be homeschooling my kids, I would have laughed. We had no homeschooling background, no plan for it, and honestly, no idea what we were doing. Yet here we are. My wife and I homeschool our two children (daughter 10, son 8). We’ve been doing this for each of them since they were in second grade (both of them did preschool in a private preschool and K and 1 in the public elementary school). Since it was a bit of a surprise for us, it also comes as a surprise to most friends and family, so I thought it would be good to write down how we came to the decision. ...

What if you could have multiple AI agents working on your codebase in parallel, each fixing bugs, simplifying code, or building features—all while you focus on the bigger picture? That’s what I’ve been trying to build. I’ve been working on automating more parts of my software development process, now that LLM coding is able to do a majority of it. This has been possible for over a year now, with less capability, but promising early signs. The first hint of it for me came from installing Cline in VS Code and using that for my first vibe coding session. Lots of others attempted improvements, from Roo to Kilo, and other VS Code plugins, but also standalone systems from the big AI developers, like Codex CLI, Gemini CLI and their web counterparts, and, of course, Claude Code. ...