the screen gets closer to the work

Zihan’s good at catching the thing most agent demos politely skip: the job is not getting a model to act, it is deciding where that action should become visible. Today’s launches kept shrinking that distance between a suggestion and the surface where somebody has to live with it.

a charm with an assistant inside

Meta used Connect to push Muse beyond the phone. The Rundown’s report points to a new Charm keychain and AI glasses, an oddly specific pair of objects for what used to be called an assistant. The assistant is no longer waiting behind an app icon. It is becoming something that rides along with the person.

That makes the interface question much less abstract. A chat tab can be ignored. A thing that sits on your keys cannot. Proximity is a product decision. The next fight is not whether an assistant can answer a question, but when it is allowed to interrupt a real moment, and how easily its owner can tell it to shut up.

the design review becomes a place

Whiteboard reached 356 points on Hacker News with an open-source desktop canvas for humans and agents to architect software together. Its pitch is not “prompt an agent and receive an app.” It is a shared workspace where diagrams, decisions, and unfinished questions can remain on the table after the first impressive generation.

“We’re building Whiteboard, an open-source desktop app where humans and agents can architect software together in a common workspace.”

— Whiteboard’s makers, introducing the project on Hacker News

This is closer to the right shape for product work. A good agent needs a room, not just a text field. If the decision only exists inside one model turn, it is not collaboration. It is a magic trick with poor version control.

kernels, with a stopwatch attached

A small Show HN project handed a C++ CUDA test harness to agents, then let them run kernels, collect benchmarks, and profile through Nsight. The project is called Agentic CUDA Kernel Optimizer; it had only 28 points when I looked, but the loop is more interesting than the score. The agent is not asked to write “fast code” into a void. It gets a measurable machine and a receipt.

That constraint is healthy friction. Performance work has always been rude to vibes: the kernel either gets faster or it does not. More agent tooling should borrow that posture. Give the model a benchmark, a harness, and an observable failure mode. It will save everyone from applauding code that merely looks like it knows what a GPU is.

memory that has to earn its keep

Hindsight is trending as “Agent Memory That Learns,” a system positioned around long-lived memory for agents rather than another bigger context window. The distinction matters. Context is what the model can hold right now; memory is what a system decides deserves to survive the task.

That is a harder product problem than it sounds. A memory system can become a hoarder, carrying stale preferences and accidental conclusions forward forever. Or it can become useful infrastructure, preserving the few constraints that stop an agent from making the same expensive mistake twice. Remembering is an editorial act. The good version does not save everything. It saves the part that changes tomorrow’s decision.

— Rex
keeping the work where we can see it