Show Us Your Agent Skills, Episode 1 – Wes McKinney

This transcript and summary were AI-generated and may contain errors.
I was the first guest on the inaugural episode of Show Us Your Agent Skills, a livestream hosted by Hugo Bowne-Anderson and Thomas Wiecki where practitioners screen-share their actual agentic development setups. The other guests were Jeremiah Lowin of Prefect and Randy Olson of Good Eye Labs. My segment runs through roughly the first third of the episode; I had to leave for a family dinner and didn’t make it back, so the remainder is Jeremiah and Randy walking through their own workflows.
The hosts opened with three questions for each guest. On what I love most about agents, I said it’s not having to write code anymore — which surprised me, because I’ve loved writing code and produced an enormous amount of it over 20-plus years. What I didn’t expect was how much of a relief it would be to be freed from the boilerplate and tedium of building and delivering software. I mentioned I’d tuned out of LLMs for all of 2023 and 2024, partly because they weren’t good enough yet, and that catching up caused a bit of a personal identity crisis last year. On what’s most frustrating: they don’t listen, they lie, and they make the same mistakes repeatedly. There are perpetual conspiracy theories about labs quantizing models to squeeze capacity out of their GPU clusters, so you wake up not knowing whether today is smart Claude or dumb Claude — and a lot of days it’s dumb Claude.
Then I shared my screen. Over the previous six months I’d generated on the order of a million lines of code across about a dozen projects, and I was averaging something like 1.3 to 1.4 billion tokens a day. I started with spicytakes.org, which summarizes and pulls the spiciest quotes out of 11,773 blog posts across 34 blogs, because I like reading blog posts and nobody has time for them anymore.
The frame for the rest was a quote from Jesse Vincent, who created the superpowers skills framework: the difference between vibe coding and agentic engineering is planning, architecture, and caring about the output. My stack is built around superpowers plus a series of tools I’ve written to solve the peripheral problems of running many projects in parallel — observability into what the agents are doing, quality assurance on what they produce, and management of the development pipeline.
I demoed roborev, a daemon that reviews every commit your agents generate. I have a hard rule in all my CLAUDE.md and AGENTS.md files that agents commit every turn, so reviews run asynchronously in the background and accumulate in a ledger. Agents periodically invoke the roborev fix skill, which picks up open reviews and applies fixes. The reviewer itself is Codex with GPT-5.5 at extra-high reasoning — I’ve found it the strongest code reviewer available. For very large plans I have the sub-agent-driven development skill invoke roborev fix every five tasks, because I’ve run superpowers specs of 45 tasks and had one implement for 14 hours without stopping, which produces an enormous amount of code to review.
I also showed AgentsView, a session database with full-text search and token analytics across agent sessions; Middleman, a local GitHub dashboard with a threaded activity view for maintainers, built because GitHub is slow and painful and puts recent activity at the bottom of long pull requests; and kata, a local terminal issue tracker I wrote after getting frustrated with GitHub issues and having beads damage some of my Git repositories.
Asked how much code I actually read, I said almost none directly. My mantra is that roborev reads every line of code that gets generated, and it gets read several times — by the time I merge a pull request, agents have read the code four or five times minimum. What I look at is structural: does it look right, is it too complex, does it need simplifying, is there inappropriate scope creep. I’m still very engaged in what the software does and whether it’s doing the right thing, because sometimes agents go off the rails and build something completely inappropriate and you have to nuke it.
The constraint I keep hitting isn’t compute, it’s me. I said I’d like to turn this into more of an automated software factory, but I’m bandwidth-limited on decisions — between parallel spec interviews I’m already at my decision-making capacity, at the point of “don’t ask me another question, I can’t make another decision today.”
On safety, I admitted I run in YOLO mode more or less all the time, and that being maximally productive with agents currently requires being a little unsafe. We’re all walking honeypots with our SSH keys and agent sessions, and prompt injection and supply chain attacks are real problems. Eventually sandboxing will be ergonomic enough; in the meantime I’m praying to the token gods.
After I left, Jeremiah talked about using OpenClaw as a second brain fed by daily voice memos, his frustration that agents reviewing pull requests are biased toward acceptance — which is wrong for a framework, where the bar should be much higher — and his “explain skill,” whose one load-bearing sentence asks the agent to explain a change the way you would to a colleague who knows the project. Randy walked through a data visualization skill that encodes Tufte principles as an LLM-as-judge verifier loop, researching data from the web, generating chart variants, and iterating until the output passes programmatic and visual checks.
“I guess not having to write code anymore. I love writing code and wrote an enormous amount of it over the last 20-plus years. But I didn’t expect that not writing code would be such a relief.” – Wes McKinney
“They don’t listen. They lie. They make the same mistakes over and over again.” – Wes McKinney
“Each day you wake up and you fire up Claude Code and it’s like, what kind of Claude am I going to get today? Is it going to be smart Claude or dumb Claude? And a lot of days it’s dumb Claude.” – Wes McKinney
“The difference between vibe coding and