WES MCKINNEY

Wes has loved writing code for decades. His favorite thing about agents? "Not having to write code anymore." In his factory, agents commit every turn, RoboRev reads every commit, fix skills drain the queue, and Wes audits structure, not lines.

▶ WATCH SEGMENT $ FACTORY WRITE-UP

1.4Btokens a day, prior week 1Mlines of code in six months $21.7Kper month at API rates 4-5Xeach commit read by agents before merge 14Hlongest unattended superpowers run 4-5projects in parallel

this episode is also a blog post: The Agentic Software Factory and a podcast: Agentic Engineering and the Lost Art of Verification

EP 01 · WES MCKINNEY · the software factory, live on stream

THE SOFTWARE FACTORY

from the agentic-software-factory write-up. fork it

"The difference between vibe coding and agentic engineering is planning, architecture, and caring about the output." Jesse Vincent's line, Wes's operating principle. Typing prompts into a terminal stops working once quality and scale matter. The factory is his answer.

The agentic-software-factory write-up captures Wes's production loop as one system: agents commit every turn, a RoboRev daemon reviews every commit in the background, review feedback accumulates in a ledger, and agents invoke a fix skill to drain it. Around that core, Superset gives every project worktree the same terminal stack, Middleman is a single pane of glass over GitHub activity, kata tracks issues locally, and AgentsView makes past agent sessions searchable.

Inside the write-up: the six rules below, a session walkthrough, anti-patterns, and a "what you need" list to assemble the stack yourself. The demo target on stream was Spicy Takes, his AI-summarized blog reader.

Wes's macOS widget showing his Claude / Codex usage split and projected API cost — Wes's macOS widget on stream: roughly three-quarters Claude Code, one-quarter Codex over 30 days, and $21,765.80/month at API rates. [00:16:14]

"Each day you wake up and you fire up Claude Code, it's what kind of Claude am I going to get today? Is it going to be smart Claude or dumb Claude?"

His sharpest frustration with agents: "they don't listen, they lie" and "they make the same mistakes over and over again." 00:07:17

HOW THE FACTORY RUNS

six rules. every timestamp opens the segment

Commit every turnA hard rule in every CLAUDE.md / AGENTS.md. Post-commit hooks give the review daemon something to fire on, and intermediate states stay inspectable. 00:15:49

Review every commit in the backgroundRoboRev runs Codex with GPT-5.5 at high reasoning on each commit, "the strongest code reviewer out there." Reviews never block generation. 00:18:42

Drain the ledger explicitlyThe RoboRev fix skill picks up open reviews, fixes them, and commits the fixes. On large plans it runs every five tasks so review debt never piles up. 00:19:37

Run parallel spec interviewsFour or five Superpowers planning sessions across projects in a day; one implementation ran fourteen hours without stopping. 00:20:36

Watch the factory, not the codeAgentsView for session search and token analytics, Middleman for cross-repo GitHub activity, kata for local issues, Superset for worktrees. 00:22:03

Audit structure, not linesWes's attention goes to complexity, scope creep, and "is it doing the right thing". Line-level reading is the agents' job. 00:27:52

The review prompt RoboRev sends to Codex, with GPT-5.5 high-reasoning configuration visible — The literal review prompt RoboRev sends to Codex: "you are a code reviewer" with the diff inlined, run at high reasoning effort. [00:18:50]

Middleman's threaded activity feed across pull requests — Middleman, Wes's local GitHub dashboard: a threaded activity feed of pushes, comments, commits, and RoboRev reviews across all projects. [00:22:17]

"I'm already at my decision-making bandwidth. I can't make any more decisions."

The ceiling on the factory isn't agent capacity. Wes wants to push toward a more automated factory, but human judgment about what the software should do is the bottleneck. 00:30:16

THE STACK

roborev, middleman & kata are wes's own builds. superpowers is jesse vincent's

RoboRevreview daemon AgentsViewsession search Middlemangithub dashboard katalocal issues Supersetworktree manager Superpowersskills framework

"We're all just walking honeypots with our agent sessions."

Wes is "YOLO mode all the time": maximum productivity over safety, mitigated by backups and keeping sensitive data out of his home directory. The field, he says, still needs ergonomic sandboxing. 00:31:28