Dan Shipper on Codex as daily driver: every agent needs a human, and AI is actually increasing demand for experts

Dan Shipper runs Every, a 30-person AI-native media and software company that has grown from four people since the GPT-3 era — while automating everything it can. That tension is the actual story he's telling right now.

Shipper says Codex has become his daily driver, and its defining feature is access to everything on his computer. He describes asking Codex to surface people he should send his latest article to, and it pulled names from his emails and text messages over the prior three months without being prompted further. The practical jump, in his framing, is from AI that's good at documents and information gathering to AI that can actually take action across your data.

“Every single person has access to an agent. Have access to as many tokens as they possibly can use. And yet, we've grown from four to almost 30 people since GPT-3. So like, what the fuck is that about? Even though AI can do expert human work, it actually increases the demand for human experts. Because what happens is you can get expert human work out of an AI, but it's all based on yesterday's competence.”

The human sandwich

Every agent needs a human. The further an agent runs from human oversight, the worse it performs — and Shipper says this holds even inside OpenAI and Anthropic, where company-wide bots are managed by dedicated teams. His colleague Kieran Klassen calls it the "human sandwich": AI collapses hours of work into minutes, but a human still frames the task at the front end and evaluates the output at the back end.

The more counterintuitive claim is that AI is increasing demand for human experts, not reducing it. Shipper's logic is that AI produces competent work based on yesterday's training data, which floods the market with output that's "kind of good, but not quite right." Human experts become more valuable because they're the ones who take that cheap baseline competence and turn it into differentiated, situation-specific work. Without them, the output is indistinguishable from everyone else's.

SaaS pocalypse is overblown

Shipper is skeptical of the idea that vibe coding kills mainstream SaaS. His estimate: 99% of people won't maintain their own apps. The harder question — raised in the conversation — is whether long-tail software gets replaced not by users building their own apps, but by agents that simply do the underlying task natively, making a dedicated app unnecessary.

Shipper's response is that chat interfaces are poor fits for many app interactions, and early Codex or ChatGPT users are effectively being trained into power users who will eventually want to buy software to solve the problems they've discovered. He's more interested in what he calls "Codex-native apps" — software designed to run inside the browser pane of an agent orchestration environment, where the agent and the user work side by side in real time. He describes spending his day inside Codex with documents, emails, and a live SaaS app all running in the same browser tab, and calls it the most powerful working environment he's used.

On layoffs and AI credit

Every has grown to 30 people despite aggressively deploying agents across engineering, writing, editing, and design. Shipper argues companies announcing AI-driven layoffs while simultaneously claiming record business performance are playing a bad hand optically — and that many of those reductions would have happened regardless of AI.

His near-term signals: new model vibe checks are coming in the next few weeks, and Every's agent product is targeting a beta launch by end of June.