tufte

Show Us Your (Agent) Skills Episode 1 - w/ Wes McKinney, Jeremiah Lowin, & Randal Olson

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 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? ​Thomas Wiecki (PyMC Labs) and Hugo Bowne-Anderson (Vanishing Gradients) are on a mission to find out. Think Excel World Championships meets Eurovision. ​First on the roster are Wes McKinney (creator of pandas, POSIT), Jeremiah Lowin (Prefect), Hilary Mason (HiddenDoor), and a few more surprise guests. #ai #aiagents #llms Our Q&A will happen on Discord so come join us there and ask questions live (# show-us-your-agent-skillz channel)! https://discord.gg/Xd7TQDuU Subscribe to our lu.ma calendar to find out about more events like this: https://luma.com/calendar/cal-8ImWFDQ3IEIxNWk Check out our Substack: https://hugobowne.substack.com/ Come build the future of Agentic Data Science with us in our upcoming course: https://vanishinggradients.short.gy/data-science-agentic Github repo for Show Us Your (Agent) Skills here: https://github.com/hugobowne/show-us-your-agent-skills 00:00 First episode: an agentic data science community sharing what actually works 05:36 Wes McKinney on relief from boilerplate, and dumb-Claude vs smart-Claude days 10:14 A million lines of code in six months: spicytakes.org and the agentic engineering stack 15:47 RoboRev: a daemon that reviews every commit, with GPT-5.5 as the strongest reviewer 20:30 Agents View, Middleman, and Kata: the rest of Wes’s local-first toolchain 27:40 Wes barely reads code; RoboRev reads it four or five times before merge 31:02 Auto mode, judgment turning into intelligence, and the YOLO sandbox tradeoff 36:08 Jeremiah Lowin on agents as a second brain, fed by morning voice memos 39:37 An open source maintainer’s guide to saying no, and FastMCP’s issue-first model 46:40 The explain skill: one sentence that changes the tenor of every review 49:03 Skills vs MCP: steering behavior vs distributing business logic 51:12 Anatomy of a skill: front matter, progressive disclosure, a polite note 56:30 Personal software and the Claude Hub power law, then Prefab and Cardboard in action 01:03:42 OpenClaw for memory, Claude and Codex desktops for parallel coding 01:07:38 Randy Olson on building at the speed of curiosity while touching grass 01:13:08 Designing the data viz skill: thin drivers, environment setup, reflect-and-improve 01:20:03 Brain, harness, skills: why minimal harnesses are winning 01:23:44 Encoding Tufte into an LLM judge, from movie deaths to marble racing 01:27:00 Live run: marriage and divorce in the US, with the verifier loop in action 01:34:32 Hugo runs the same skill on Secretariat’s 1973 Kentucky Derby record 01:36:40 Generator-evaluator workflows, and how vigilant to be when agents game the judge 01:43:17 Wrap up and upcoming guests