PT

SERV is proud to be among the founding partners of IC, collaborating with leading protocols like NEAR, MetaMask, OKX, Nansen, BNB Chain. Bringing a shared trust layer to the agent economy, projected to drive $5 trillion in commerce by 2030. Entire agentic stack in one flow 👇.

OpenAI's new GPT-5.6 $1 model (Luna) just beat the $5 flagship (Sol) on our agentic benchmark. With SERV Reasoning, all three models performed better, while Luna outscored every configuration tested. Thanks to SERV, failure rates dropped by up to 42.7%.* That's the thesis SERV is built on: Reliability in AI is the real product. Large-scale AI adoption is waiting on the layer that makes AI dependable enough for production. Standout findings: - Luna + SERV tops the table at 1/5 of Sol's price. - These tiers behave less like mini/nano distillations and more like independent takes on one architecture with distinct post-training. - Luna behaves like a smaller model RL-trained hard for agentic instruction-following and steerability, which would explain why it takes the biggest lift from SERV. *i.e. the relative drop in failure rate: Luna went from 12.89% failed tasks to 7.39%. More results and insights coming soon.

I’m that guy who wears a Halloween costume only one person gets. So it’s no surprise I decided to build an AI tool for a very niche and very human job: supervising therapists in training. I’m a psychiatrist, not an engineer, so I made every mistake you can think of, but I did learn (the hard way). One of the hardest lessons came early. I gave my tool to a resident I work with and the tool said some offensive shit—I was there when it happened. Sounded smart, but mistimed, miscalibrated, and intrusive. What to do? I built a second agent to watch the first. Problem was, the second agent had the same brain, was made of the same substrate, and suffered from the same stupidity. So then what? Infinite recursion? A conga line of accommodation? What’s my point? Well, our boys just wrote a deep dive on Shadow Agents, and I smiled reading it, because it’s my watcher agent but made by Unc-led geniuses. Their version: every output is checked against expectations and if it doesn’t conform, the model goes back and tries again. Two years of watching mistakes in production allowed the team to guide the agents uniquely in an eloquent distillation. For banking and robotics among plenty of other categories, it just works. And it works because a bank can write down what “right” means—what rules, in what order. A robot arm instruction is either well formed or it isn’t. Write the rules down completely and you can train a checker to enforce them. The loop ends inside the machine, legitimately. That’s what makes agent fleets possible in places where “usually works” gets you fired. My problem is different. In supervision, there are many right ways to approach things. There’s no rulebook to adjudicate—not everything is declarative. There’s no parameter for judgment hard earned through human experience. So my loop can’t end reflexively. It ends with me, a human, reading the logs and updating the code. That’s the dividing line imo, and most folks have no idea there’s a difference. Tasks where “right” can be written down to the last detail can be verified at machine speed and scaled the fuck up. Tasks where human judgment and taste remain quintessentially important require a human looped in. OpenServ calls it supervision as infrastructure. I’ve spent almost six months on the other kind of supervision, trying to help people grow. Same word, doing two jobs that have almost nothing to do with each other. Know which one your problem is before you go full agentic army. Otherwise you might just get a real scare. Boo.