Prefect

On a team call today, evaluating tools, I said: if your tool doesn't have a good MCP server, I'm not going to use your platform anymore. I thought that was a controversial take. Everyone on the call agreed. SaaS isn't dead. The way we interact with it is becoming obsolete, fast. As more of our work moves into agents, the tools that adopt that shift are going to be the ones still standing. The ones that don't will get phased out. It also means your organization is about to be running a lot more MCP servers than you realize. Agents pulling from dozens of them, all hitting different systems, all touching different data. You're going to need governance for that. Who's allowed to call what, where data is flowing, how to keep it secure. That's what we're building Prefect Horizon for.

h@ trick works at the drop of an @ we know you wear many @s h@s off, here's where it's @ 🚀 Launch week begins tomorrow. Schedule in the comments ☟

My AE team is out-adopting most engineering teams on AI. Not Prefect's engineers (they're elite). Engineering teams broadly, the function historically synonymous with early AI adoption. Our AEs don't use AI to do the job, but to eliminate everything that isn't the job. The oversimplified job of an AE is three things: (1) Show up prepared and ready to add value, (2) build credibility and trust with real humans, and (3) provide the right information at the right time. So they codified our best practices into automated workflows with built-in prompts, skills, tools, and the right triggers and schedules. The low leverage work just runs, and our AEs show up with the judgment and relationship skills that AI can't touch. - Pre-call research and prep, delivered before the call - Follow-up emails drafted with a human in the loop - CRM hygiene and enrichment - Risk and red flag assessment in active evaluations - Champion enablement materials built around the buyer's actual evaluation All of it was eating rep time and capacity that could have gone toward supporting more buying teams. Fokke Dekker and Trent Broderick built this playbook themselves, proactively, because they understood the risk of getting left behind, and the upside of taking charge. We're still iterating, but I'm impressed with where the team has landed, and curious how far this goes. If your team is doing something similar, or think we're missing something worth adding, drop it in the comments. Powered by Anthropic products, MCP servers from our favorite GTM tools, and our own MCP platform, Horizon.

What's your stack for building MCP apps? I've been building web apps a long time. The stack keeps changing. Django. React. Next.js. Python. JavaScript. TypeScript. The interfaces shift. The motive doesn't. Reduce friction for the user with the tools you have. Lately I've been obsessed with a different shape. The MCP app stack. Still an app. Db, services, UI. But the interface shifted. The agent client is the new doorway. Claude Desktop. Claude Code. ChatGPT. Copilot. Working at Prefect, I've had a front-row seat to what's being built in-house. After a year of building MCP servers, I've settled on a Prefect-fueled stack for my own MCP apps. This is what it looks like on a project I'm working on ↓

Geert-Jan Van den Bogaerde just shared "There's an xkcd for that" with me - bringing definitive, argument-ending punchlines directly to your agent. And of course, built with FastMCP + Prefab. https://xkcdforthat.com/

Jeremiah Lowin joined Thomas Wiecki, PhD and Hugo Bowne-Anderson for the first episode of Show Us Your (Agent) Skills. If you missed the stream, Jeremiah talked about agents as a second brain, custom memory, and building software shaped around the way he works. They also made a great intro for him, featuring an 8-bit Jeremiah slicing through robots with a glowing keyboard sword ⚔️ Link to the full conversation in the comments ☟

"There is a dopamine hit you get when you see your sub flow fan out into a bunch of things and then come together at the end." Nithin Krishnamurthi, Software Engineer at Snorkel AI, walks through how Snorkel replaced a tangled mess of Redis queues with Prefect as their chief orchestrator. His PyAI Conf 2026 talk covers the custom worker architecture that brought observability and sanity to a wildly diverse compute stack spanning on-prem model training, RL evals, and 24-hour packaging scripts. Full talk here: https://lnkd.in/g87ykw4W

Pydantic double dog dares you