Wes McKinney on AI Agents, The Mythical Agent Month, and His Wild AI Coding Setup
TL;DR
Wes McKinney went from muting 'LLMs' on Twitter to building his life around coding agents — after dismissing AI through 2023 and much of 2024, he says Claude Code and especially Opus 4.5 finally crossed the line from toy demos to software he could actually trust enough to build with.
His core claim is that agent-generated code is still 'hot garbage' unless multiple other agents review it — McKinney built RoboRev to run adversarial reviews on every commit, saying he now does 4,000-5,000 reviews per month and often has code read 10-20 times before shipping.
The 'mythical agent month' is his warning that people confuse working demos with real products — echoing Fred Brooks' The Mythical Man-Month, he argues the jump from software that appears to work to a viable, maintainable product is still enormous, and coding agents are best at the Potemkin facade, not the hard operational guts.
He’s using agents to reclaim personal software from decaying SaaS platforms like Gmail — his MessageVault project pulled roughly 50 GB and 2 million emails out of Gmail in about 12 hours via OAuth, storing them locally with SQLite and DuckDB so he can search and analyze his own data outside Google's tools.
Go has become his favorite 'language for agents' because cycle time matters more than human ergonomics — after finding Python too slow for build/test loops and Rust too slow to compile, he says Go's sub-10-second builds make it ideal for agent workflows that constantly tweak, build, and test.
His setup is basically an AI coding air-traffic-control tower — using Kitty, multiple terminal panes, Git worktrees, AgentsView, RoboRev, and now superset.sh, he runs six to eight Claude/Codex sessions at once, with agents beeping for attention like 'checking the cake in the oven.'
The Breakdown
From AI skeptic to all-in builder
McKinney opens by admitting he basically ignored AI through 2023 and much of 2024 — he literally muted "LLMs" on Twitter. Then came what he calls a stretch of existential dread about the future of software engineering, followed by a turn after seeing Steve Yegge post about Claude Code and realizing newer models like Opus 4.5 had crossed a real threshold.
Coding agents made software feel fun again
Once he started using Claude Code on Positron at Posit, he says the big unlock was not just writing code faster, but avoiding the parts of development he hated: sysadmin, DevOps, CI/CD, packaging, notarization. His energy in this section is pure relief — agents filled in 20 years of skill gaps and turned the work that used to make him "shrivel up a little bit" into background labor.
Why raw Claude output still isn't safe to ship
McKinney is blunt: if you're committing agent output directly, you're shipping slop. That realization led him to build RoboRev, a system where every agent turn creates a commit, every commit triggers reviews, and those reviews get fed back into the same session so bugs and edge cases are fixed while context is still fresh.
MessageVault and the war on Gmail decay
His personal-productivity obsession starts with email, and the rant lands because it's so specific: 20 years in Gmail, terrible search, daily frustration. So he built MessageVault, used OAuth to export around 50 GB and 2 million emails in roughly 12 hours, indexed everything locally with SQLite and DuckDB, and now plans to keep Gmail nearly empty so he never has to suffer its search UX again.
The 'mythical agent month' and Potemkin software
This is the conceptual heart of the interview. Borrowing from Fred Brooks' estimate that the gap between working software and a real product is about 9x, McKinney argues that vibe coding makes people overconfident because agents are excellent at building the "porcelain" or "Potemkin facade" while the invisible maintenance, scaling, security, and operability work still dominates real products.
Why Go beats Python and Rust for agent workflows
His reasoning is practical, not ideological: if humans aren't hand-writing the code, readability matters less than build/test speed. Python felt too slow in execution, Rust compiled too slowly for iterative agent loops, and Go hit the sweet spot with single-digit-second builds, making it ideal for terminal tools, utilities, and backend services where agents constantly modify, build, and test.
Inside the wild multi-agent setup
McKinney's daily environment sounds like a mission control room: Kitty terminal in three columns, multiple splits, six to eight Claude or Codex sessions visible, per-project review queues, and agents configured to beep whenever they stop. He compares it to checking "the cake in the oven," and says he now uses superset.sh to manage Git worktrees more cleanly while RoboRev handles the heavy-duty review pipeline in the background.
Where agents break: big codebases, false AGI confidence, and software hoarding
He pushes back on AGI-style claims by pointing to his own experience: getting from zero to 50,000 lines of code is easy, but somewhere around 100,000 lines, agent effectiveness drops sharply, and million-line codebases like VS Code need "a treasure map" plus strict lane boundaries. His closing vision is both exciting and slightly chaotic — a future where everyone becomes a software hoarder with 10, 50, maybe hundreds of personal apps, while the real moat shifts from coding itself to taste, architecture, prompting, and knowing how to steer agents through complexity.