|
AI Builders Digest
Friday, July 10, 2026
|
|
The agent stack is clicking together, as Vercel CEO Guillermo Rauch put it this week. But a field report from an actual company using AI agents all day suggests "clicking together" and "working well for humans" are two different things. The gap between a smooth technical demo and a functional team culture is turning out to be significant.
|
|
---
|
|
01
|
When everyone talks to their AI all day, nobody talks to each other
|
|
|
Zara Zhang, a VC at GGV, shared a striking data point from a founder she interviewed: the company bought OpenAI's Codex Max subscriptions for every employee, and now most work happens through those individual AI sessions. The side effect nobody planned for: meetings evaporated, collaboration dropped, and team culture took a hit. Everyone got more productive in isolation. The company got less coherent as a whole.
|
Why it matters: Your org chart assumes information flows between people. When every employee optimizes their own AI workflow independently, the institutional knowledge that usually spreads through hallway conversations and Slack threads stops moving. You end up with a team of very productive individuals who don't actually function as a team. The productivity metrics will look great until something goes wrong and nobody knows what anyone else was working on.
|
|
Source →
|
|
---
|
|
02
|
The quiet reason AI startups won't admit they're using Chinese models
|
|
|
Shawn Wang, who runs Latent Space, made a pointed observation about the agent infrastructure space: most labs quietly use Chinese models under the hood but won't say so publicly because they're selling to government and defense clients. He highlighted one company, Cog, that did the work to make it viable: building evals for multilingual propaganda and censorship tendencies, correcting for them in post-training, and serving at 1,000 tokens per second cheaply.
|
Why it matters: If you're evaluating AI vendors for anything sensitive, "what models are you running in production" is now a compliance question, not just a technical one. The companies doing the sanitization work deserve credit, but the ones staying quiet about their model stack are creating a due diligence problem for their customers.
|
|
Source →
|
|
---
|
|
03
|
Together AI launches reserved capacity for open models
|
|
|
Together AI introduced Provisioned Throughput: reserved inference capacity for frontier open models including MiniMax M3 and GLM-5.2, with token-based pricing and a 99% uptime SLA. They're claiming up to 90% lower cost than proprietary APIs, with no GPU infrastructure to manage.
|
Why it matters: The open-model cost advantage over GPT-4-class APIs has been real but unreliable — you get cheaper tokens when capacity is available, and painful slowdowns when it isn't. A reserved-capacity product with an SLA changes the calculus for production workloads. Teams that ruled out open models because of reliability concerns now have a reason to revisit that call.
|
|
Source →
|
|
---
|
|
|
|
**Box CEO Aaron Levie flags Grok 4.5 as worth watching for enterprise use** — Levie notes that the latest wave of models, Grok 4.5 included, are getting meaningfully better at legal, healthcare, and professional services work, where domain complexity previously made AI outputs unreliable.
|
|
Source →
|
|
**Guillermo Rauch posts a vague-but-enthusiastic agent stack demo** — The Vercel CEO shared a clip of "all the pieces of the agent stack clicking together" with no specifics on what he's actually building. Worth watching if you follow Vercel's product direction, thin on substance today.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|