This is a reserve digest for signals that are weaker than standalone posts, but useful as a map of what is being assembled around AI right now.
Agent infrastructure. Veridect, marked sponsored in the batch, sells a policy layer before agent actions: allow, escalate, or block. The category is right, but claims about 3-second decisions and tamper-proof logging need independent verification. Ramp Router offers an OpenAI-compatible endpoint and claims up to 30 percent cost reduction; the important questions are routing policy and quality tradeoffs. OpenShip is useful as a self-hosted deployment platform, but the AI angle is still weak on its own.
Interfaces. Notion AI browser use is interesting as a signal of a browser/action layer inside a productivity app, but public technical detail is thin. Rivet connects a local agent with design references, so it belongs near the shared-document theme. Linear Loops is better read as an update to Linear-as-control-plane than as a separate thesis. “12-factor companies” is a useful frame: own context, rent intelligence, but it still needs an implementation surface.
Video and developer workflows. Lucy 2.5 continues the “video as interactive medium” line: real-time editing, reference control, and 30 FPS. Excalidraw plus Cursor for database diagrams is a small reproducible workflow. Plan -> Codex -> Kimi review shows a multi-model review loop, but model consensus does not prove quality. The finite UI elements thread is a useful reminder: a coding agent should assemble UI from an approved component registry instead of inventing every button again.
New Runtime read: policy layers, shared surfaces, video tools, and workflow recipes are all forming around AI at the same time. Not every signal deserves its own post, but together they show the shift from “AI answers” to “AI is embedded in the production system”.