Player not loading? Watch on YouTube
Wes McKinney, Jeremiah Lowin and Randy Olson explain how they use agent skills for software development and data visualization. The discussion covers planning, review and the human decisions that remain in an AI agent workflow.
McKinney describes a stack centered on Superpowers, with RoboRev reviewing commits in the background and a fix skill addressing outstanding feedback. His review configuration uses Codex with GPT-5.5 and extra-high reasoning, which he rates as the strongest reviewer he has tried. Agents View provides session search and analytics; Middleman is a local GitHub dashboard, while Kata tracks issues in the terminal. These local tools do not establish that the underlying models run offline.
Lowin explains why he still reviews framework contributions manually. He uses an explain skill to get context about a change rather than a line-by-line recap, and distinguishes skills that steer behavior from MCP servers that distribute business logic. OpenClaw's customized memory supports work spread over time; Claude and Codex desktop apps handle his parallel coding sessions.
Olson demonstrates a data visualization skill that researches sources, tries chart variants and checks results with code and an LLM judge informed by Tufte's principles. The marriage and divorce example still has an overlapping annotation. The guests discuss evaluator variation and the need for deterministic checks when a failure must be prevented.
Several guests present tools they created or maintain. The hosts also promote their agentic data science course.
