OpenAI's Q1 Results: $5.7B in Revenue, but $3.7B Burned
According to The Decoder, OpenAI's revenue for Q1 2026 reached **$5.7 billion**, roughly tripling year over year. Meanwhile, expenses for the quarter came to about **$3.7 billion**, which also nearly tripled from the previous year.
Particularly striking is that stock-based compensation alone consumed **over $2.3 billion**. That said, with roughly **$73 billion** in cash on hand, the company reportedly has no pressing need to raise new funds for the time being. Still, if the pricing war with Anthropic intensifies, the situation could change quickly.
While the rapid revenue growth reflects the expansion of the AI market, the results also laid bare a structure in which the war for talent and infrastructure investment continue to squeeze profits.
The AI Agent "Approval Rush" — Permission Gates Are Wearing Humans Down
Two interesting articles were published on Zenn. Both deal with permission-management problems that surface when actually operating AI agents.
The first introduces the concept of the "approval rush." When you use an AI agent whose tool executions are gated behind permission, the agent constantly asks, "May I run this?" Even after you approve one request, more keep coming — and when you deny something, the agent comes back asking again with slightly different wording. Users gradually wear down under this barrage. It's a paradox: approval gates built for safety end up causing "consent fatigue" instead.
The second article reports a hands-on experience of running an MCP server with tightened permissions. The author installed freee's MCP server with all tools permitted, only to discover that the AI — asked simply to register an expense — had on its own accessed the account master and the vendor list. No actual harm was done, but the author reflects on having proceeded with an "as long as it works" mindset without reviewing the tool list, and argues for the importance of designing minimal permission scopes.
The lesson shared by both articles is clear: the safety of AI agents should be ensured by the **precision of permission design**, not by the number of approvals. Reckless approval requests erode human attention and end up weakening security instead.
Vercel Eve: The "Agent = Directory" Design Philosophy
Eve, the agent framework Vercel released this week, is drawing attention. According to an analysis article on Zenn, Eve's defining trait is captured in a single official sentence:
> An agent is a directory.
When deploying agents to production, the biggest time sink is not model selection or prompt tuning but the "plumbing": durable execution, sandboxing, human approval flows, tracing, and evaluation. Rebuilding all of this from scratch every time is a major burden.
Eve tries to solve this by hiding the plumbing behind a directory convention. Configuration files and workflow definitions are expressed as a directory structure, letting agent behavior be managed on a convention basis. As a result, developers can focus on their agent's logic.
What makes Eve interesting is that while traditional frameworks took a "describe your agent in code" approach, Eve presents a paradigm shift: "declare your agent in a directory structure."
Why AI Agents Waste Tokens on Windows
Another notable Zenn article analyzes the problems that arise when using GitHub Copilot's agent mode on Windows. Contrary to the expectation that products from the same Microsoft stable should work well together, the agent's terminal operations reportedly fail frequently on Windows.
The cause lies in the behavioral differences between PowerShell 5.1 and Bash. The agent tries to chain commands with `&&` and gets rejected, and `ls -la` throws an error. Commands that assume Bash behavior simply don't work in PowerShell, so the agent ends up retrying the same operation over and over. These retries lead directly to wasted tokens.
Upgrading to PowerShell 7 or leveraging WSL can improve things to some extent, but fundamentally you need to be aware of the gap between the shell environment an AI agent assumes and the actual execution environment.
Wrap-Up
This edition covered a broad range, from the macro-level flow of money into the AI industry to the micro-level practical challenges of operating agents. OpenAI's earnings show how hot the market is, while the approval rush and token waste on Windows are issues you face immediately when using AI agents in real work. The evolution of frameworks like Vercel Eve can be seen as a step toward solving these plumbing problems.
---