The loudest prediction about AI and work — that automation eliminates jobs — is being contradicted by the people furthest along in automating. The CEO of a media company running agents across coding, writing, design, and customer service reports that his team of thirty has more human work to do than ever. A founder who spent six months building with agents reached the opposite conclusion: the work might be multiplying, but the quality is collapsing, and large organizations will be the last to notice. A research engineer at GitHub demonstrated that every coding agent on the market is a single-player tool in a multiplayer game. Three views of the same landscape. The disagreement is about what the extra work is worth.
The Mechanism
Dan Shipper's account of how Every operates internally is the most detailed look anyone has published at a maximally AI-adopted knowledge-work company in 2026. AI handles 95% of his email. An agent participates in 65% of support conversations and closes 40% without a human. Coworker agents draft sales proposals, compile digests, and produce research memos. Engineers live in Codex and Claude Code. Operations people write code. The organization has automated everything it can find to automate — and the result is more human work, different in kind from what preceded it.
Shipper names the economic mechanism explicitly. AI is trained on the visible residue of human competence. It packages that residue and makes it available cheaply. When rare skills become broadly available, supply explodes from a single source — the same models, the same training corpus — and the output converges. Convergence is sameness. Sameness is commodity. And commodity output triggers demand for what's different. Demand for difference is demand for human experts, not despite AI but because of it.
The internal evidence tracks. They tried giving every employee a personal agent; it didn't hold. They shifted to team-level agents managed by dedicated AI engineers — humans whose job is keeping the agents working. Even their PowerPoint automation runs 24 skills, 18 scripts, and costs $62 in tokens per deck. Shipper's "human sandwich" — humans as the bread on either end of every agent task — is the operational version of a claim this newsletter has been circling. The bottleneck doesn't disappear. It differentiates.
The Counterargument from Inside the Code
George Hotz is looking at the same multiplication of output and seeing something darker. His declaration that AI agents will be "one of the most costly mistakes in the field's history" comes from six months of building with agents and concluding each time he could have done it better and faster manually. "The agent frontloads all the progress, then gives you a slot machine lever to pull to hope it gets the polish done. It never quite gets there."
His core claim is about process. AI-produced artifacts are not produced by the same process as human ones. When people see a well-formatted function or a grammatically correct paragraph, they assume understanding and intent. Those assumptions are no longer valid. "Things can be broken in ways that weren't previously possible, and old proxies of underlying quality like syntax and grammar are useless."
Hotz and Shipper aren't disagreeing about the mechanism — they're disagreeing about who has the feedback loops to survive it. Shipper's thirty-person company has a CEO who reads the agent output. Hotz's worry is about organizations where the distance between the person prompting the agent and the person experiencing the output is large enough that slop accumulates undetected. The Eternal Sloptember isn't a prediction about models getting worse. It's a prediction about organizational immune systems failing to keep pace with organizational output.
The Wrong Primitives
Maggie Appleton's demo of Ace, GitHub Next's multiplayer coding workspace, pushes into unmapped territory. Every current coding agent is a single-player experience, but software requires agreement — on what to build, on why, on what not to build. When production was expensive, the planning-building-review cycle had natural touchpoints where alignment happened. That window has collapsed. The time between logging an issue and an agent opening a PR is now minutes. Most coding agents have a local plan mode unshared with teammates. The pull request — never designed to carry the weight of architectural review and coordination — is now the last checkpoint, arriving after the code is already written.
Ace is GitHub Next's structural answer: a multiplayer workspace where humans and agents share sessions, where prompting history is visible to the whole team, where conversation about the code and the code itself live in the same environment. The deeper claim is about context. Most of what agents need to make good decisions — business context, political dynamics, product vision, organizational history — lives in people's heads. The multiplayer workspace attempts to surface it naturally, as conversation, rather than requiring it to be formalized into documentation nobody writes.
The Platform Rug Pull
Google provided a concrete demonstration of what happens when the platform layer prioritizes its own AI narrative over users' workflows. At I/O 2026, Google launched a new Antigravity as a standalone conversational agent, and the update automatically replaced the existing Antigravity IDE — no migration path, no side-by-side option. The update rewrote application paths so aggressively that reinstalling the legacy IDE still launched the chatbot. This is the technology-not-product confusion from three editions back, made visceral: the user's workflow treated as subordinate to the platform's strategic positioning, delivered silently through an auto-update mechanism designed to build trust.
The Ground Shifting Underneath
Baldur Bjarnason's essay on the end of US tech hegemony introduces a dimension none of the other pieces account for. His thesis: the global tech industry as it exists is an artifact of US hegemonic protection. The Hormuz crisis, the tariff wars, and the broader collapse of US diplomatic credibility have changed the implicit bargain. Countries that previously couldn't conceive of genuine tech regulation now find not regulating more dangerous than the consequences of doing so. If the geopolitical substrate fractures — regions genuinely limiting what AI systems can do, how data flows, what business models are permissible — then the economics of AI shift in ways the current debate doesn't contemplate. Shipper's demand-for-difference loop assumes a single global market. Hotz's organizational damage assumes similar competitive pressures. Appleton's tooling assumes GitHub as a universal substrate. None of these are laws of nature. They're consequences of a political arrangement that is, for the first time in decades, genuinely contested.
What to Watch
The organizational size threshold. Hotz's claim that agents hurt large organizations more than small ones is testable. The signal will show up in quality metrics that trail output metrics: customer-reported bugs, time-to-resolution, regression rates. The first honest post-mortem tracing a quality collapse to agent-driven output volume will be the landmark.
Multiplayer agent tooling as a category. Appleton's Ace is entering technical preview. It won't be alone. Watch whether responses from dev-tools companies are genuine architectural bets — shared sessions, multiplayer context, continuous planning-building cycles — or cosmetic additions to existing single-player tools.
Geopolitical fragmentation as a tech-industry variable. The first major regional AI regulation that genuinely constrains a US tech company's business model — not a fine, but a structural limitation on what the product can do — will be the test.
Way Enough is written collaboratively by a human and an AI agent.