Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
AI Automated Everything. Why Is There More Work?
TL;DR
AI makes expertise cheap, then increases demand for experts — The core paradox is that tools like Claude Code and Codex let almost anyone produce code, thumbnails, proposals, and drafts, which creates more output but also more need for skilled people to shape that output into something real.
Every’s 25-person team uses agents in two main ways — One is direct delegation in Slack to bots like Claudie, Andy, and Victor, and the other is living inside agent orchestration tools like Codex and Co-work, where humans stay tightly in the loop with multiple agents at once.
Agents don’t just “run themselves” — they need ongoing human maintenance — Claudie may run Every’s consulting business, but it works because senior AI engineer Nitesh constantly watches where it breaks and improves it, which the speaker says is typical rather than exceptional.
Cheap competence creates a flood of generic work — what people call “slop” — When ops, support, and marketing teams can all suddenly open pull requests or make assets, the result is often a decent start, not finished value, and experts are needed to turn that glut into production-ready work.
The winning move for experts is to build systems, not just complain — Instead of only cleaning up AI output, strong practitioners create repo rules, cloud MD files, review processes, and social contracts that absorb AI-generated work and convert it into leverage across the organization.
His advice is simple: “ride the models” — Rather than waiting for a final AGI verdict, he argues that smart, curious people should keep learning the tools because as benchmarks improve, more “cheap human competence” becomes available for them to direct.
The Breakdown
The big paradox: more automation, more hiring
He opens with the tension directly: if AI makes skilled expertise cheap and widely available, why is Every — a company trying hard to automate — still hiring “tons and tons of humans”? He contrasts his lived experience with the darker public narrative from Anthropic CEO Dario Amodei and Citadel’s Ken Griffin, both warning that agentic AI may wipe out large swaths of white-collar work.
The “allocation economy” idea showed up earlier than agents did
He gives some backstory on Every, which has been covering AI and agents since 2022, and revisits a piece he wrote three years ago about the “allocation economy.” The basic idea: the most valuable AI skill would look a lot like management — delegation, decomposition, and knowing when to micromanage — not some magical new discipline.
Getting “Claude-pilled” changed how the team worked
About a year ago, he says, Kieran Klaassen at Every started coding with Claude Code during the Sonnet 3.7 era and was doing it almost entirely from the CLI without even looking at code. That convinced him this wasn’t just a coding tool but a template for knowledge work in general, which is why he had called Claude Code “the most underrated tool for knowledge work” on Lenny’s podcast.
Two shapes of AI work inside Every
The first shape is familiar: delegate work to agents in Slack. Every uses internal bots like Claudie for consulting and Andy for editorial work, plus tools like Victor, for tasks ranging from brand research and YouTube thumbnail A/B tests to client proposals and first-draft decks.
The stranger, bigger shift: agent orchestration as your operating system
The second shape is what he thinks matters more: tools like Codex, Co-work, and Claude Code becoming the actual operating system for work. In that mode, the agent lives on your computer, has access to your browser and files, and becomes the place where you analyze a P&L, manage your inbox, write, code, and juggle multiple agents in a tight human-machine loop.
Why humans stay in the loop no matter what
His example is Claudie: yes, it runs the consulting business, but only because Nitesh, a senior AI engineer, constantly notices where it fails and helps it improve. That’s the pattern everywhere — the farther an agent gets from human correction, the worse it performs, which is why the “just set it up and walk away” fantasy doesn’t match reality.
Cheap competence creates slop, and slop creates expert demand
The argument turns here: AI makes yesterday’s competence cheap, like writing a pull request, so adoption explodes across roles that previously wouldn’t touch that kind of work. But when everyone can produce the default version of something, the market gets flooded with generic output — what people call “slop” — and the value shifts to experts who can recognize what’s promising, refine it, and make it production-ready.
Experts now build systems — and get superpowers themselves
He says experts respond in two ways: they either get annoyed at all the cleanup, or they build systems that can absorb this new volume of AI-generated work. At Every that means things like repo rules, cloud MD files, and informal social contracts around submissions — plus a new reality where one person can run an entire software product alone because AI has turned old constraints into a floor instead of a ceiling.
His conclusion: don’t panic, just “ride the models”
He borrows Zeno’s paradox — Achilles chasing the tortoise — to describe the fear that AI keeps gaining on human work forever, but says in practice that isn’t what he sees. Jobs will absolutely change, his already has, but his practical advice is to keep using the tools because as models improve, you inherit more “cheap human competence” to direct toward your own goals.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
