|
AI Builders Digest
Wednesday, November 4, 2026
|
|
There's a quiet argument happening in AI infrastructure right now, and it's worth paying attention to. Everyone building agents is discovering the same problem: the environments those agents run in weren't designed for them. Yesterday we covered OpenAI rebuilding Git from scratch because Codex broke it. Today, Vercel's CEO is making the same point from a different angle. The boring plumbing is becoming the whole story.
|
|
---
|
|
01
|
Vercel's CEO says stateless sandboxes are the wrong way to run AI agents
|
|
|
Guillermo Rauch, CEO of Vercel, made a pointed argument this week: AI agents are only useful if they carry their full context with them, including identity, skills, repositories, and memory. The current industry default of "stateless sandboxes," where an agent spins up fresh with no memory of prior work, is measuring the wrong thing. Rauch noted that Vercel's sandbox infrastructure now spans 19 global regions and is seeing double-digit weekly growth in usage.
|
Why it matters: If your company is evaluating AI coding tools based on how fast they spin up a blank environment, you're benchmarking the wrong thing. The agents that will actually ship useful work are the ones that remember what they did yesterday. The ones that don't are expensive autocomplete.
|
|
Source →
|
|
---
|
|
02
|
Google DeepMind's drug discovery agent is searching chemical space for novel compounds
|
|
|
Google DeepMind published details on Google Antigravity, an agentic system built for drug discovery. The system can search compound libraries and propose novel "rigidified macrocycles," a class of molecules that are notoriously difficult to design by hand because of how many structural variations need to be evaluated simultaneously.
|
Why it matters: Drug discovery is one of the first fields where AI agents have a defensible advantage over humans, not because they're smarter, but because the search space is too large for any team to cover manually. If Antigravity's compound proposals hold up in the lab, this is a preview of what "AI finds the drug" actually looks like in practice, not as a press release, but as a chemistry workflow.
|
|
Source →
|
|
---
|
|
03
|
Cisco is running OpenAI models as its first line of cyber defense
|
|
|
OpenAI published a case study on how Cisco is integrating its models into network security operations. The pitch: combine Cisco's network visibility with OpenAI's pattern recognition to respond to threats faster than a human security team can triage them.
|
Why it matters: "Machine speed" defense has been a marketing phrase in cybersecurity for a decade. The difference now is that the models are good enough that "automated response" doesn't just mean "block the IP." It means reasoning about whether an anomaly is a threat, which is a qualitatively different capability. If this works as described, it shrinks the window attackers have between getting in and getting caught.
|
|
Source →
|
|
---
|
|
|
|
**Replit's Amjad Masad on agent security:** Masad posted that AI agents should run in organization-controlled environments with minimal access to only what each task requires. This is less a product announcement and more a design principle, but it's worth bookmarking. Every team that gives their agents broad permissions today will regret it specifically.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|