Aaron Levie has been on a roll this week, and today's post connects directly to yesterday's enterprise AI thread in a way worth pausing on. The throughline: coding agents work because code is falsifiable in real time. Almost nothing else is.
01
Why AI coding agents work, and why that advantage is hard to copy
Box CEO Aaron Levie posted a thread making a point that sounds obvious but has real consequences: code is special for AI agents because you can test it immediately. Write a function, run it, see if it works. Most other professional work doesn't have that feedback loop. When an AI agent drafts a contract, executes a trade, or writes a sales pitch, you don't know if it worked until the deal closes or falls through, which might be weeks later. ---
Why it matters: The coding agent boom is not a preview of the agent-for-everything future. It's a preview of the agent-for-testable-work future. If you're a vendor pitching "AI agents for [legal/finance/HR]," you're selling into domains where the error surface is invisible until it's very expensive. Buyers should be asking not just "can it do the task" but "how fast does it know it failed."
Vercel plugs AgentMail directly into its agent install flow
Vercel CEO Guillermo Rauch posted about an integration with AgentMail: developers can now run `vercel install agentmail` and their AI agent gets email capabilities automatically, no signup, no separate billing. It's a small announcement, but it's doing something specific. Yesterday Rauch was showing agents that deploy and A/B test. Today they also have an inbox. ---
Why it matters: Vercel is quietly turning itself into the place where AI agents live, not just where websites get deployed. Each integration that removes a signup step makes it slightly easier to build an agent that does something real in the world. The billing consolidation is also worth noting: when agents start racking up third-party service costs, knowing they all show up on one invoice matters.
OpenAI's Codex is apparently burning through free credits at a rate that's becoming a conversation
Thibault Sottiaux, whose bio places him at OpenAI, posted asking whether to reset the ChatGPT Work and Codex usage limits again, with usage approaching 9 million. The post got 1,600+ replies, which suggests this landed with people who have opinions about it. ---
Why it matters: OpenAI is managing a genuine tension. Codex and ChatGPT Work are free or heavily subsidized right now, usage is growing fast enough that someone inside the company is publicly asking when to pull the throttle, and the community clearly has strong feelings about the answer. If you're building anything that depends on current Codex access levels, the free ride has a clock on it.
**Peter Yang posts a link with zero context** — The post is literally "Holy shit a time traveler or cope?" with a link and no further explanation. Without knowing what the link goes to, there's nothing here to report. Dropped. **Thariq shares a Pokemon team breakdown built as an artifact** — Thariq posted his first public artifact: a breakdown of his Mega Sceptile competitive Pokemon team, with a note that he'll open source the underlying tool if there's interest. It's a fun demo of the artifact format in a non-work context, and the open source offer is genuine if you're curious about the tooling.