Your next user is not a human

Your Next Power User Doesn't Have a Face

Aaron Levie said it plainly last week: “As agents become the biggest users of software, all software has to be available in a headless fashion.” Translation: every product that requires a human clicking through a screen to operate is going to need a version that works without one, because the next power user isn't a person. That shift is already underway.

Cloudflare and Stripe shipped a joint protocol last week that lets AI agents create their own Cloudflare accounts, register domains, and deploy live applications without a human touching anything. Stripe verifies the agent and caps its spending at $100/month per provider. Supabase, Hugging Face, Twilio, and AgentMail are already on the protocol. An agent can now spin up infrastructure, pay for it, and ship a working app to the internet on its own. The human's role is setting the budget and saying go.

Salesforce made the same bet at a much larger scale. At TrailblazerDX, they launched “Headless 360”, exposing their entire platform so agents like Claude Code and Codex can operate on Salesforce directly, no browser required. When the world's largest enterprise CRM rebuilds its architecture around the assumption that the next user is an agent, that's a strategic decision about which side of the transition the company wants to be on.

The pattern is consistent: the infrastructure layer is rebuilding itself around agent consumers. Whether you're buying or selling software, the evaluation question is already shifting from “is the UI good” to “is the API good,” and any product that still requires a human in the loop to operate is accruing real strategic debt against that transition.

The talent is following the same signal. Former CTOs and senior builders from Workday, Box, Adept, You.com, and Instagram's orbit are leaving executive seats to take hands-on technical roles at Anthropic. When software becomes downstream of models, the smartest operators want to be closer to the model layer than the management layer above it. Running 300 engineers inside an old stack matters less now than getting your hands on the tooling and product surface that's actually shaping where things go.

Our read: companies rebuilding their architecture for agent consumers right now are going to own the next distribution layer. The ones still optimizing dashboards and onboarding flows for humans are polishing a product whose best customer is about to be a bot.

The Difference Between Vibe Coding and What Comes Next

Andrej Karpathy co-founded OpenAI, got Autopilot working at Tesla, and coined the term vibe coding. Last week at AI Engineer World's Fair, he said he's never felt more behind as a programmer, then spent 30 minutes explaining why that's the point.

The distinction he's drawing: vibe coding raises the floor. Everyone can build now, and that changes everything. But agentic engineering is about preserving the quality bar of professional software while going dramatically faster. You're still responsible for your code, you still can't ship vulnerabilities, and the discipline is coordinating these spiky, stochastic, extremely powerful agents without sacrificing what used to make good engineering good. Our read: the 10x engineer conversation is over. The people who are good at coordinating agents are pulling away from everyone else, and not by 10x.

1. Invest in your setup like it's your editor. Engineers used to obsess over vim configs and VS Code extensions. Now it's Claude Code, Codex, or OpenClaw, and the ones who are investing serious time into their agent environment, their project context files, their workflow are pulling away from everyone else. The setup compounds on every task the same way a good dev environment always has.
2. Write specs, not code. Work with your agent to design a detailed spec before generating anything. You own the architecture, the data model, the “these have to be unique user IDs” decisions. The agents fill in the blanks. Karpathy's cautionary tale: his agent tried to match Stripe email addresses to Google email addresses to associate purchases with user accounts, instead of using a persistent user ID. The kind of mistake that seems obvious to a human but falls outside what the agent was optimizing for.
3. Forget the API details, keep the fundamentals. Karpathy says he's already forgotten whether it's keep_dims or keep_dim, dim or axis, reshape or permute. He doesn't need to remember anymore. What he still needs: understanding that there's an underlying tensor view vs. a copy, and that copying memory around is expensive. The layer above the API is yours. The API itself belongs to the agent.
4. Hire for this. If you're still giving programming puzzles in interviews, you're hiring for the old paradigm. Karpathy's version: give someone a big project, have them build a Twitter clone, make it secure, deploy it, then throw 10 agents at it to try to break what they built. That tells you more about how someone coordinates and directs agents than any whiteboard exercise ever could.

Roon captured the cultural version of all this in a tweet that got 700K views last week: “people are walking around with their laptops slightly ajar to keep their agents running.” The image is a little dramatic, but the behavior is real. The people who are good at agentic engineering have agents running constantly, testing and building and exploring. The laptop is open because the work doesn't stop when you walk away from the desk.

Try it: the full Karpathy talk is on YouTube, 30 minutes, worth every one.

Also worth your time

The Pentagon just signed AI deals with 7 Big Tech firms

OpenAI, Google, Microsoft, Amazon, Nvidia, SpaceX, and Reflection AI will deploy AI on classified military networks. Anthropic was notably excluded after refusing to remove safety guardrails around autonomous weapons and mass surveillance. Worth watching how this reshapes the relationship between AI labs and the defense establishment.

Anthropic and OpenAI just stood up AI services arms

Anthropic announced a $1.05B joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to help private-equity-backed businesses integrate Claude into their operations. OpenAI is launching a parallel structure with its own partners. The data point that explains the move: customers are reportedly spending $8 to $12 on services for every $1 they spend on Anthropic software. The labs are following the actual revenue.

97% deployed agents, 29% seeing ROI

A new Writer survey on enterprise AI adoption puts numbers on what everyone suspects: nearly every company has agents running, almost nobody is seeing real returns yet. 54% of C-suite say AI adoption is “tearing their company apart.” 29% of employees admit to actively sabotaging AI strategy. The deployment gap closed. The execution gap hasn't.

$300 billion in one quarter

Q1 2026 venture funding shattered every record. Four of the five largest rounds in history closed in 90 days: OpenAI ($122B), Anthropic ($30B), xAI ($20B), Waymo ($16B). That's $188B to four companies, or 65% of all global venture investment for the quarter.

Grok 4.3 shipped quietly

xAI cut API pricing, improved speed, and added stronger tool use without the usual spectacle. The gap is still real though: Grok scores 53 on Artificial Analysis while GPT-5.5 sits at 60 and Claude Opus 4.7 at 57. Useful for writing and customer support. For hard reasoning and multi-step execution, still chasing the pack.

Novo Nordisk went all-in on OpenAI

The world's largest pharma company by market cap is integrating OpenAI across drug discovery, clinical trials, manufacturing, and commercial ops, with full deployment by end of 2026. Eli Lilly signed a competing deal with Insilico Medicine in March. Big Pharma's AI arms race is real and moving faster than most people in the industry expected.

See you next Tuesday. Sven