Prefect

A lightning talk from @VstormCommunity : @antonikozelski, CEO and Co-founder, has shipped 30+ production agent solutions across industries. In his PyAI Conf 2026 talk, he walks through three open source tools Vstorm is releasing back to the community, including a @pydantic Logfire AI assistant that lets anyone (even non-technical folks) interrogate their agent systems through natural language in a Chrome extension. 📺 Full talk here: youtube.com/watch?v=GBpOsV…

OpenAI's goblins came from training their Nerdy personality. Radhika Gulati's (@radhigulati) goblin reference come from her being a nerd. See: her October post on killing the context goblin with an MCP server. Nerd to nerd, see link below 👇

"Why are we still creating feature pipelines by hand?" Chang She @changhiskhan, CEO and Co-founder of @lancedb and original co-author of the pandas project, sees a world where software engineers have stopped writing most of their code manually but data teams are still debugging distributed pipeline failures at 2 AM and watching their OpenAI bills spike over the weekend. In his PyAI Conf 2026 talk, Chang breaks down the ten biggest pain points of feature engineering for AI and makes the case that our data infrastructure was built for the last generation. Full talk here: youtube.com/watch?v=dv4vZQ…

Prefect Learns is live. Wayne Grigsby, our Senior Sales Engineer, takes you from your first flow all the way to atomic transactions, dynamic runtime flows and concurrent execution. FastMCP is in there too if you're building agents that need real tools. Easiest track to hardest, pick where you start. All free. prefect.io/learn

"Unlike code, data is non-local by default. If you change something, there's a bunch of production systems that depend on what you change. And unfortunately, we do not have git." @GreCo_CiRo, Co-founder and CEO at @Bauplan_labs, explains why AI coding assistants hit a wall when they try to touch production data and what to do about it. Every pipeline run is an isolated branch, merges are atomic, and failed runs never reach production. His PyAI Conf 2026 talk includes a live demo of an AI agent building and deploying a data pipeline with full git-like version control for your data lake. 📺 Full talk here: youtu.be/WnqvGcO2-nA

The "MCP is dead" argument might just be the gap between what servers can do and what clients support. @jlowin met Sam Morrow @SamMorrowDrums at Dev Summit, @AAAzzam caught up with him at AI Engineer London — he's the author of the most popular MCP server on the planet, and he's living that gap daily. New ep on Spotify, link below 👇

"If someone hands you a Jupyter notebook, you might die a little inside if they said, get this in production." @DylanMadisetti, Software Engineer at @marimo_io, at PyAI Conf 2026. Dylan shows why Python notebooks don't have to be a production nightmare. marimo builds a compute graph behind every notebook, eliminates hidden state, enables reactive widgets, and stores everything as pure Python you can lint, test, and import like any other module. Full talk here: youtu.be/xDY3APX9qp8

Happy Friday to Us. Our team discovered the draw function in Claude Design. The Prefect Horizon One Pager got updated with an essential new testimonial in about 5 minutes.

"If you actually go put an agent in front of a consumer, if they don't see something happening in literally two seconds, most of them will just navigate away and do something else." Bradley Axen, Tech Lead for AI at @Square and maintainer of @goose_oss, breaks down the three architectural patterns that actually matter when building with MCP, skills, and MCP Apps: progressive disclosure, output optimization, and human-in-the-loop done right. His PyAI Conf 2026 talk includes a live demo of Goose ordering a Uji Matcha Latte from a real Square MCP server. Full talk here: youtube.com/watch?v=2RrF8k…

What are people at the top of the game building with AI agents and how are they doing it? ​Are they Claudemaxxing with 8 terminals open at once? Or adversarially testing @AnthropicAI Opus 4.7 generated code with @OpenAI Codex? Do they define suites and swarms of sub-agents or use AGENTS.md and agent skills? ​What do they love about building with agents? What do they hate? What tips and tricks do they use to supercharge their workflows? @twiecki and I are on a mission to find out. Think Excel World Championships meets Eurovision. ​First on the roster are @wesmckinn (creator of pandas), @jlowin (Prefect), @hmason (HiddenDoor), and a few more surprise guests. ​Register to join live or get the recording afterwards (link in comments).

Anyway, here's a video of our engineer Zach promoting run club for our annual offsite. We're starting to think he has a TikTok account that's doing numbers and we just don't know about it 👀

The forecast for MCP summer is hot. We just got back from the MCP Dev Summit and AI Engineering Europe so this one's a little different. Instead of breaking down a feature or a release, we're talking about the McPeople. The ones building MCP, the ones using it, the ones showing up to argue GitHub issues with Jeremiah in person (note: this is cool). @AAAzzam and @jlowin recap their favorite talks, conversations, and escalator run-ins from a week across two conferences: @idosal1 demoing an agent orchestrator inside World of Warcraft. @liadyosef and Ido's MCP Apps keynote rewiring how we think about what's coming. Max Isbey giving what he CLAIMS was his first talk. And @digitarald from VS Code, who @pamelafox told Jeremiah he couldn't miss. These tags and more on the full episode below.

If you spent the whole weekend prompting Claude Code and wondering whether any of that context survives past the session, then this is the lightning talk from PyAI Conf 2026 for you. @indragie Karunaratne built Claudescope at @sentry because he noticed a problem: when agents write the code and other agents review it, nobody knows what the human actually intended. The PR becomes a 10,000-line artifact with no story behind it. Full talk at youtube.com/watch?v=IPBIB7…

FastMCP 3.2 is out with full MCP Apps support. Create interactive tools, replace your team's dashboards, and use your MCP server as a complete backend for your app. You can even let agents generate UIs on the fly:

"If you've built a Turing complete YAML, maybe just use a Turing complete language." Ethan Rosenthal , Senior ML Engineer at @runwayml, at PyAI Conf 2026. He's spent ten years watching every ML project follow the same path: script, CLI, config file, Turing-complete YAML nightmare. Weights and Biases sold for almost $2B and one of their core features was tracking arguments to functions. Ethan built Confingy to solve that problem in pure Python, and he's open sourcing it in this talk. 📺 Full talk on YouTube: youtube.com/watch?v=ElR9He…