|
AI Builders Digest
Thursday, July 9, 2026
|
|
Aaron Levie has now spent two days in a row explaining why AI agents keep failing inside big companies, and the pattern he's describing has nothing to do with the models. The bottleneck is organizational. That's a polite way of saying the problem is people, and specifically the way companies are structured to resist exactly the kind of cross-functional work that agents require.
|
|
---
|
|
01
|
Enterprise AI agents are stalling out on org charts, not model quality
|
|
|
Box CEO Aaron Levie just finished a road trip through meetings with dozens of enterprise IT leaders, and the diagnosis he came back with is worth sitting with. The models aren't the problem. The operating model is. Companies built their teams in siloes, but agents work best when they're tied to a full process that cuts across those siloes. Nobody owns that process. Nobody has budget for it. So agents get deployed inside one team, deliver partial value, and the project stalls.
|
Why it matters: If you're the person inside a large company trying to drive AI adoption, this is your actual obstacle. You're not losing to a competitor with a better model. You're losing to an org chart that was designed before anyone thought about AI agents. The companies that figure out centralized agent governance first won't just be more productive, they'll make it structurally harder for the rest to catch up.
|
|
Source →
|
|
---
|
|
02
|
Vercel's eve agent framework gets filesystem-based tool extensibility
|
|
|
Vercel CEO Guillermo Rauch showed off how eve handles tool integration: drop a `tools/github.ts` file into your project, export a `createGithubTools()` function, and your agent has GitHub powers. No configuration menus, no plugin marketplace. Just the filesystem.
|
Why it matters: Yesterday we covered eve shipping built-in evals. Today it's the extensibility model. Rauch is making a consistent bet that agent frameworks should feel like web frameworks, where conventions replace configuration. If that bet lands, the developer who already knows how to build a Next.js app will be able to build an agent without learning a new mental model.
|
|
Source →
|
|
---
|
|
03
|
Microsoft ships an open-source charting language built for AI agents
|
|
|
Microsoft Research published Flint, an open-source visualization language designed to solve a specific problem: AI agents are bad at writing chart specifications because the popular libraries (Vega-Lite, ECharts, Chart.js) require verbose, low-level configs that models get wrong constantly. Flint sits in between. You write a compact, human-readable spec. Flint's compiler handles the design decisions: scales, colors, label spacing, whether a y-axis should start at zero. One spec compiles to any of the three backends. The project ships with a Model Context Protocol server so agents can generate and render charts directly in chat or coding environments.
|
Why it matters: Anyone building an AI assistant that produces reports, dashboards, or data summaries has probably hit this wall. The chart looks broken, the model apologizes and tries again, the user loses confidence in the whole product. Flint is a narrow fix for a specific annoyance, but it's the kind of narrow fix that quietly ends up everywhere once enough developers discover it.
|
|
Source →
|
|
---
|
|
04
|
Together AI launches reserved inference capacity for open models at 90% off proprietary pricing
|
|
|
Together AI is now offering Provisioned Throughput: reserved inference capacity for frontier open models including MiniMax M3 and GLM-5.2, with a 99% uptime SLA and token-based pricing. The pitch is up to 90% lower cost than proprietary APIs, without requiring customers to manage their own GPU infrastructure.
|
Why it matters: The case for self-hosting open models has always been cost, but the hidden cost is the engineering time to keep it running reliably. Together is positioning this as the option that gets you the savings without the ops burden. If that math holds, the "run it yourself" argument gets a lot harder to make to a CFO.
|
|
Source →
|
|
---
|
|
05
|
Enterprise data and evals is a massive opportunity, per Madhu Guru
|
|
|
Madhu Guru, who has worked in AI models for enterprise, vouched for the team at Mercor and added that the opportunity around data and evals for enterprise is, in his words, "massssive."
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|