Two stories in today's payload are making the same point from different angles: getting AI agents to work reliably is mostly a context problem, and the solutions people are reaching for look less like science fiction and more like a shared folder on a network drive.
01
The file system is the new API for AI agents
Box CEO Aaron Levie put into words what a lot of builders have been circling around: the reason most agents fail isn't the model, it's that the agent can't find what it needs to do the job. His argument is that shared file systems, where both the human and the agent can read and write the same plans, notes, and drafts, are what actually make agents useful in practice. "What they need is a working set: plans, notes, task lists, policies, drafts." ---
Why it matters: If Levie is right, the companies winning at enterprise AI won't necessarily have the best models. They'll have the best-organized data. Your company's messy SharePoint folder isn't just a productivity problem anymore; it's an AI capability problem. The teams cleaning up their internal documentation right now are quietly building an advantage.
Vercel CEO: markdown is the new programming language
Guillermo Rauch, CEO of Vercel, shared the structure of a minimal agent built entirely from markdown files: an `instructions.md` and a skills folder, deployable in one command. His take is that we're at the most accessible moment in programming history, where defining software behavior means writing plain text that humans can read and agents can execute. ---
Why it matters: This connects directly to yesterday's Claude Code artifacts story. When your agent's "code" is a markdown file anyone on the team can edit, the bottleneck shifts from engineering to whoever writes the clearest instructions. That's a real org chart change, and most companies haven't felt it yet.
Someone used Claude Code to attempt to decode a 3,500-year-old undeciphered script
Boris Cherny shared a project where Claude Code was applied to Linear A, the Bronze Age writing system from Crete that has stumped linguists for over a century. The work is up for peer review, and Cherny's hedging appropriately ("Hope this holds up!"), but the attempt itself is notable. ---
Why it matters: Linear A has resisted decipherment because there aren't enough surviving examples and no bilingual reference text. If AI can make meaningful progress here, it suggests coding assistants are quietly becoming research tools for problems nobody thought to apply them to yet. Keep an eye on whether this peer review holds.
Latent Space co-founder Swyx dropped a prediction with no elaboration: Anthropic hits $2T at IPO. The linked article isn't visible in the post, so the reasoning behind the number isn't clear. The post is thin on substance, but the number is worth noting. Anthropic's last known private valuation was in the $60-80B range. A $2T IPO would put it in Apple territory. Either there's a serious bull case being written somewhere, or this is Twitter math. ---
PSPDFKit founder Peter Steinberger posted about a new team member who "speaks both developer and agents." One sentence, no details. Nothing here to report beyond a hiring announcement.