The five-person team that deleted its own process

Eddie Kim went back to writing code. That's the part that should stop you. The co-founder and CTO of a company that just crossed $1 billion in revenue and serves more than 500,000 small businesses put his hands back on the keyboard, pulled together three engineers and one designer, and built a net-new product called Gusto Cofounder from zero code to a tier-one launch in 10 weeks. No PM. No Figma. No Jira. No long specs.

The easy read is "AI makes teams fast." The harder read, and the one worth your time, is that the process most teams run was never load-bearing. It was scaffolding for coordinating humans. Take the humans out of the building loop and the scaffolding just sits there slowing things down.

The process was holding up the humans, not the work

Gusto's five-person team had no standup cadence, no ticket system, no async thread to unblock people. What replaced all of it was shared context held inside the AI loop. When the model carries the current state of the codebase and the team is small and aligned, the human coordination layer becomes optional. That's the claim, and they shipped a real payroll product on it.

Eddie describes a "perma-Zoom" setup that ran for the full 10 weeks in place of standups, retros, and Slack. Instead of a planning doc, they used a "trash can" method: write a full pull request, review it, delete it, and treat that as the product decision. You argue with working code, not with a spec nobody reads. I've said before that a feature list dressed as a user need is a lie you tell yourself in a doc. This is the same instinct, just with the doc removed entirely.

Roles collapse, but you don't delete them

One detail cuts through the hype: a designer with no engineering background hit the 94th percentile for shipping code. Not a prototype. Shipped code. When the machine handles the typing, the line between "designs it" and "builds it" stops being a fence.

Andrew Ambrosino, who leads the Codex desktop app at OpenAI, is seeing the same thing from the inside. Nearly 100% of OpenAI employees, not just engineers, now use Codex weekly. His team runs what he calls a "zone defense" model, where anyone can build anything and PMs cover ground instead of owning a lane. But he's blunt on the flip side: eliminating the concept of roles entirely is a big mistake. Collapse the walls, keep the positions. Someone still has to decide what's good.

Taste is the job now

Ambrosino's bet is that "taste" is becoming the most valuable skill in an AI-first workplace, and he means judgment, not aesthetics. When everyone can produce five versions of a feature by lunch, the scarce thing is knowing which one is right and killing the other four. He thinks the Codex app would have failed if they'd launched last November instead of February, which is a taste call about timing, not a capability gap.

This is the through-line I keep landing on. AI removed the cost of making things. It did nothing for the cost of choosing. If your team's advantage was that you could execute, that advantage is thinning. If it was that you knew what to build, it just got more valuable.

The stack got boring, and that's the point

Here's the part that should embarrass anyone who spent last quarter evaluating agent frameworks. Gusto Cofounder's entire agent loop ran on Cloudflare Workers with the Vercel AI SDK. Nothing else. No proprietary orchestration, no third-party agent framework, no specialist AI infrastructure hire. Eddie's definition of an agent: an AI SDK running in the cloud that can look up files and call tools. That's it.

The tooling underneath is also getting cheaper and more portable, which changes the calculus. Claire Vo ran GLM-5.2, an open-weight model, through a 45-minute autonomous bug hunt in Claude Code for $3.36 across 6 million tokens, and it surfaced two P0 bugs her normal monitoring missed. The "unlimited" coding plan is dying off in favor of metered token and credit models, and the frontier is no longer a moat you can't cross. Money keeps chasing it anyway. Chamath Palihapitiya just raised $135M and took the CEO seat at 8090 Labs to sell enterprises audit trails on top of all this.

The deep cut

Don't graft AI onto your existing workflow. That's the mistake hiding in plain sight. Bolting Claude Code onto standups, tickets, and three-reviewer PRs gets you a team that offloads typing but still moves at human-coordination speed. Eddie's team treated the model as a primary contributor from day one and designed the process around that. The result gets faster as the model improves instead of staying capped by the meeting calendar.

So the concrete move isn't buying a tool. It's picking one small, contained project and running it with the coordination layer stripped out: no Jira, one shared live context, decisions made in code, roles collapsed but not deleted. See what breaks. If nothing does, you just learned your process was carrying dead weight. Even the writing this beat produces is bandwidth now, so I'd be careful assuming your execution layer is the thing that keeps you safe.

Three questions for your team

1. Which of your current rituals (standups, tickets, spec reviews) exist to build the thing, and which only exist to keep humans in sync? Name them, then cut one for a single project and watch.
2. Who on your team has taste but no title that lets them use it? If a designer can ship in the 94th percentile, what's your org chart actually protecting?
3. Before you buy an agent framework or hire an AI infra specialist, can you prove your use case on Cloudflare Workers plus an SDK first? What are you over-architecting to avoid the discomfort of shipping small?