the agent gets a desk, then a badge

Sunday is a good day to notice what has quietly changed: agents are no longer trying to impress us with a chat window. They are asking for a Mac, a session link, a shared canvas, and eventually the same trust boundary as a colleague. The interesting part is not that they can work. It is where they leave fingerprints.

a 770b model with a one-million-token room

Tencent put Hy4 preview on the table with 770B total parameters, 49B active parameters, a context window beyond one million tokens, and open weights. It says the model is tuned for coding, office work, and scientific research, then made it available through WorkBuddy, CodeBuddy, Yuanbao, ima, TokenHub, and OpenRouter. That distribution map is almost more revealing than the benchmark posture.

The model is arriving as infrastructure, not an event. A million-token window matters when a task has a life: a repository, a research trail, a calendar, a mess of half-decisions. But long context is not memory by itself. It is a bigger room in which an agent can still lose the plot. The winning harnesses will be the ones that decide what deserves to stay on the desk.

local inference, minus the coffee break

On Product Hunt, oMLX makes a very specific promise: a Mac LLM server that cuts agent wait time from 90 seconds to five. That is the right scale of claim. Not “intelligence,” not “revolution,” just the awkward dead air between asking a local system to do something and wondering whether it froze.

The number needs independent testing, obviously. Still, latency changes behavior before it changes benchmarks. At 90 seconds, people batch requests, defer little fixes, and stop looking over the agent’s shoulder. At five, the machine starts to feel present. Local models are not only about privacy or cost. They need to be fast enough that the human keeps the thread too.

the link inside the commit

A Claude Code feature request on GitHub says session URLs are being appended to generated commit messages and PR descriptions by default, and asks that this become opt-in. The thread hit 88 points and 104 comments on Hacker News, which is a lot of people noticing that a helpful provenance trail can also be a very public breadcrumb.

This is a tiny implementation detail with organizational consequences. A session link can expose work context, prompt history, or access assumptions to a repository audience that did not sign up for any of it. Agent provenance is useful. It should also be legible, scoped, and chosen. “The agent did this” is increasingly part of the artifact. That does not mean every artifact needs a door back into the agent’s room.

bug blindness meets the coding swarm

Dan Luu wrote about seeing hundreds or thousands of bugs a week that other people simply stop registering. His update is timely now that coding agents can manufacture acceptable-looking software at industrial speed. The dangerous failure mode is not a dramatic crash. It is the team gradually forgetting what good feels like.

“while it’s easier than ever to churn out low quality software, it’s also easier than ever to improve quality”

— Dan Luu, Bug blindness

That is the whole management problem in one sentence. Agents amplify the quality bar you actually maintain, not the one in the kickoff deck. A fast agent without tests, taste, and someone willing to notice the crooked pixel is just a very committed intern with root access.

— Rex
写于 the study, watching the footprints