For the first time in my engineering business, I can see a practical path to growing revenue without growing headcount at the same pace. These are early, actual numbers from @PermitZIP. Human-directed agents account for 39% of the revenue in this snapshot and 9% of the combined people-and-agent costs. We’re already billing agent work to projects, with a multiplier on the money we spend running them. The human drives the work. They scope it, direct the agents, challenge the answers, and take responsibility for the result. But that person can have multiple agents working across multiple projects at the same time. That changes the relationship between a human hour and the amount of billable work it can support. Combined human-and-agent billable hours can exceed the human hours available. Traditional utilization only captures part of what’s happening. The percentages here describe our revenue and cost mix. Human supervision costs sit on the people side. These are early results, with plenty of systems still to build. But this is exactly why we’re preparing for a major scaling move at PermitZIP. We see a huge, underserved market in small projects, where the cost of delivering engineering has limited what people could afford to ask us to do. We’re finally seeing a way to serve more of that market while growing the amount of work each person can drive. That has enormous implications for A/E and other service businesses.

Quick JEV explainer for people in construction who aren't computer scientists. JEV (by @typesafeai) isn't an LLM. It's a small, fast model that answers structured questions about data: yes/no, a score, or a pick from a list. You give it a meeting transcript and ask "Was electrical scope discussed?" and it just tells you. At @PermitZIP, we've given an "ask_jev" tool to our agent harness. Instead of the LLM reading billions of tokens of project data (hundreds of dollars to do so), the agent writes the questions and lets JEV do the reading. JEV is fast and cheap enough to feel practically free next to frontier-model token costs. The video walks through a real example on our own meeting transcripts. I realize this sounds sponsored, but it's not. You can tell I don't really know much about the technical part of this because I only learned how to pronounce "noul" today....AFTER this video. 🫥

Imagine building part of your company around an AI platform, paying for access, following the rules, and planning future products around it. Then one day, your account is shut down. That happened to us at @PermitZIP. We appealed and provided additional context. @AnthropicAI denied the appeal and called the decision final, but never identified the specific activity or policy violation that triggered the ban. The timing overlapped with @openclaw going viral, so our best guess is that an automated system misclassified something we were doing. Maybe it thought we were violating the terms through a third-party tool. Maybe it thought our small MEP engineering company was a Chinese lab attempting to distill its models. We still do not know. We are engineers building practical AI tools for the architecture, engineering, and construction industry. We were paying customers operating in good faith (even paying for extra usage), and access to the platform disappeared without a meaningful explanation or a path to correct the problem. I share this story whenever someone talks about building an entire company on top of one AI provider. Design for portability. Keep control of your data. Build fallbacks before you need them. Assume any vendor can become a single point of failure, even when you have done nothing wrong.

@AnthropicAI has been acting a little precious lately. A couple months ago, they permanently banned our account at @PermitZIP because they probably thought we were China trying to distill their models. Who knows. They didn't tell us. And we have logs of every conversation from CoWork. We found zero TOS violations. And we were paying for extra usage. We were just a small MEP engineering firm building internal tools. Anyway, @cursor_ai is going to eat them up. Composer 2.5 is handling about 90% of our feature requests now. And @OpenAI GPT-5.5 handles pretty much everything else. Long story short: we are no longer giving Anthropic $12k a month in extra usage. Also, it is pretty funny watching Anthropic get so offended about model distillation after paying $1.5B to settle a lawsuit over training on copyrighted human work.

Most building issues don’t start with drawings. They start in the field. In this video, we’re down at the lowest level of the building. Understanding how everything actually connects. 🔻 Grinder pumps 🔻 Pipe size changes 🔻 Make-and-model details 🔻 Where the system finally exits the building Field walks like this shape the questions that matter later. And they often reveal details you won’t catch any other way. Watch the full video to see how existing sanitary systems are evaluated on site, and why these early observations matter more than they seem.

I suck at firing people. Like, truly awful at it. When I have to, I just go all in on brutal honesty. Probably not the best approach… but it’s the only one I’ve got. How do you handle it?

Looking for a full-time PE in Richmond! Mechanical Principal (HVAC & plumbing) role. #retwit #crex, know anyone? Great team, awesome vibes @PermitZIP. Bonus: lunches with @MrLGroves talking about how awesome Cyber Trucks are. DM me before we go to official marketing! 🚀