This week’s AI theme:
Agentic infrastructure matters: memory, tools, evals, sandboxes, credentials, approvals, and logs.
The model is still the engine, but the surrounding stack is where a lot of the product value is moving.
The companies that win will make agents useful inside real workflows, not just impressive inside a chat box.
News
Vercel Ship 2026 was all about agents
Vercel laid out an Agent Stack: models, gateway, workflows, sandboxed execution, channels, and secure integrations.
This is the right framing. Agents do not just need better prompts. They need production infrastructure.
The big shift: Vercel wants to be where agents are built, hosted, connected, and monitored.
Eve is a new open-source agent framework from Vercel.
The pitch is very clean: an agent is a directory.
Instructions and skills live in Markdown, tools live in TypeScript, and the framework wires up durability, sandboxes, channels, and deployment.
This looks promising because it makes agents feel more like apps: structured, inspectable, versioned, and shippable.
This is the big open-source model story of the week.
GLM-5.2 is an open-weight Chinese model focused on long-horizon coding and agentic tasks.
Early benchmark chatter says it is competitive with the latest closed models from Claude and OpenAI.
If open models keep getting this good, they will put real pressure on usage, margins, and lock-in for the big labs.
OpenAI is acquiring Ona for Codex
OpenAI’s framing: Codex needs secure, persistent cloud environments where agents can keep working over hours or days.
That is exactly where coding agents are going. They need a trusted workspace, not just a local terminal session.
The agent wars are becoming infrastructure wars.
Tweets Worth Saving
Amy Deng used AI to investigate a medical mystery
Amy is an AI researcher and brain tumor patient who used models to help crack her mystery fatigue faster than her PCP could.
The important part was not “ask ChatGPT what I have.”
It was using AI to organize symptoms, pull together context, ask better questions, and find evidence hiding in plain sight.
This is one of the clearest examples of AI as a patient-side reasoning tool. Not a doctor replacement, but a major upgrade to personal health investigation.
Rahul from Ramp on AI engineering
Really good thoughts from someone building AI inside an actual high-stakes company.
The part that stuck with me: AI code is not automatically correct code, especially as problem complexity and diff size grow.
For AI engineering, small diffs, clear ownership, evals, and review loops matter more than ever.
This is the sober version of the agent hype: agents can ship a lot, but production still demands taste and judgment.
Built
Robinghood agentic trading system
This is the kind of agent work I am most interested in: not a generic chatbot, but a system around a specific workflow.
The hard part is the surrounding product layer: what context the agent sees, what tools it can use, how it reasons through steps, and how the human reviews the output.
More and more, the winning agent products look like software systems with AI inside them, not AI demos with software around them.
Takeaway
Models are getting cheaper and more competitive.
Open-source is putting pressure on the frontier labs.
Agent frameworks are becoming real developer infrastructure.
The most useful AI products will be built around context, workflow, trust, and review.
The next question is not “can the model do it?” It is “can the system make the model reliable enough to use?”