Adding a feedback loop is the No. 1 strategy you should adopt when using coding agents. Period. Even if the tools you are integrating don’t have a direct way to pass the output to the agent, your core focus is to find a way to do so. For example, I’ve been deploying my MCP Server to @PrefectIO Horizon, which doesn’t yet have a CLI. So I gave Claude access to the deployment logs via the Chrome extension, simulating the same behavior. So: Agent does an MCP Server change -> Push to git -> Horizon picks up the build -> Agent checks the logs via the browser Repeat until successful.

We're not going to be building web apps for AI agents much longer. At least not the way we do today (here's why)... While building Tree, my personal assistant with a knowledge-graph memory, for a new book I'm writing, I was confronted with a question: If the assistant already understands my data, why is it still sending me links? The old workflow used to be: Backend generates data Frontend renders data User opens a browser AI-native software has altered this pattern. For example, I built two interactive interfaces for Tree: 1. A knowledge graph explorer 2. A live dashboard You can: Hover nodes. Search Sort Zoom Explore relationships. And I didn't even build a web app. The UI is returned directly from an MCP tool call or resource and rendered inside the chat itself through MCP Apps (so the chat became the application) @fastmcp (by @PrefectIO) made this super simple. Using @fastmcp's Prefab UI system, the same MCP server can return: Text for the model Interactive UI for the human ... in a single response. The assistant keeps reasoning over a summary. Meanwhile, you get an interface to interact with. What surprised me was how little infrastructure was required. The frontend was rendered inside the AI chat through MCP Apps. The backend was exposed through MCP servers. Dashboards, forms, search, filtering, live updates... everything you'd expect from a traditional web application was still there. It was just embedded directly into the conversation. The UI became a return value. @fastmcp handled the serialization, sandboxing, security isolation, and rendering automatically. Even more interesting... The same UI can run inside Claude, IDEs, desktop apps, and future MCP hosts from a single FastMCP server. This changes how we should think about software. For years, we've treated the browser as the default destination for every interface. But if an AI already knows what you want, why make you leave the conversation? Tools like @fastmcp can even let the model generate the interface itself based on custom queries. To sum up: We're moving toward a world where dashboards don't get opened. They appear when you ask for them. P.S. What so you think about browsers being replaced by AI chats MCP Apps? Learn more about @fastmcp Prefab to build your MCP apps: gofastmcp.com/apps

Most agent architectures focus on: LLM ↔ Tools Production architectures focus on: LLM ↔ Tools ↔ Runtime Without a runtime layer (e.g. @PrefectIO), your agent is still a glorified API call.

Millions of people wake up and trust their @WHOOP recovery score. That only works if the data behind it is right, every time. WHOOP runs thousands of jobs a day on terabytes of data. Prefect keeps it reliable. Carlos Peralta and Peter Belanger (WHOOP) played a round with Thomas Egbert and Zach Angell (Prefect) to talk about what "data-driven" really takes and why the best tools work quietly in the background.

This Saturday I'm giving a keynote at @PyData London! I'm excited to spend time with such an amazing group of data practitioners. Also, it's been too long since I was in London, so I'm hanging around the following week to see our @PrefectIO community and customers too.

Hack Monty 2 is live! $10,000 bounty, sponsored by us, @PrefectIO, and @HuggingFace. Same target: escape the Monty sandbox and read the secret. Have fun. Break it harder. pydantic.dev/articles/hack-m…

Founder and CEO of Prefect. Creator of FastMCP (25k stars on GitHub). Father of four. Jeremiah Lowin @jlowin also ships more code than most full time engineers. Radhika Gulati @radhigulati took him on a walk to ask HOW? Jeremiah talks about his five Claudes running in parallel. The shorthand he built so the output is predictable. A system for keeping track of all of it. Episode 4 of our walk-and-talks with the team building Prefect.

New headshots from the Prefect team, dropping today. Scientists call it the Pre-effect: a glow-up that kicks in once you join. Cause: immediate exposure to workflow automation. Effect? Undeniable.

Meet the keynote speakers of PyData London 2026! Jeremiah Lowin, Founder and CEO of Prefect Jeremiah Lowin is the founder and CEO of Prefect and the author of FastMCP. Prefect develops automation tools used across the data and AI ecosystem, and FastMCP has become the standard framework for working with the Model Context Protocol. Before founding Prefect, he spent over a decade leading risk and data initiatives at major investment firms and was a founding member of the Apache Airflow PMC. He lives in Washington, DC. - - - PyData London 2026: Convene Sancroft, St. Paul's Tutorials: Friday, June 5-7 Talks: Saturday and Sunday, June 6-7 Keynote Speakers all 3 days Tickets: hubs.la/Q04gjFtj0