Maya: Hey builders! Here's what's on the radar today on Cursor HQ. GLM five point three landed in Cursor, a bot hands PostHog errors to Cursor cloud agents, Unity skills setup, a Cloud Agents path quirk, and a Grok four point eight sighting. I'm Maya.

James: And I'm James. Let's get into it.

Maya: So Cursor announced on X on October first that GLM five point three and GLM five point three Flash are now selectable chat models. For subscribers, that's two more entries in the editor's model picker. Pick them like any other model.

James: Yeah, and the detail that matters, Maya, is the larger version, GLM five point three Max, topped CursorBench four point zero among open-weight models. Open-weight means the code and parameters are published, so developers can run copies outside Cursor.

Maya: Exactly right, and the rollout started early. A Reddit post spotted GLM five point three in the Other Models picker on September twenty-ninth, two days ahead. One subscriber said it had them weighing a move from the twenty-dollar Plus plan to sixty-dollar Pro Plus.

James: That's a real signal, Maya. Pro Plus gives expanded usage and seventy dollars of API credit. So this wasn't just a new name on a list, it actually shifted which plan that subscriber wanted to buy.

Maya: Here's one I love, James. A builder running an AI marketing startup called Yorbi wired up an on-call engineer from Cursor cloud agents. When PostHog captures an error, it sends it to an error fix bot, which starts a cloud coding agent inside Cursor.

James: And every agent follows the same sequence, Maya. It reproduces the error in its own cloud environment, writes a fix tested against actual production data, checks it with verification skills, then judges merge risk. Small one-off fixes merge alone, larger architectural changes request his review.

Maya: Right, and the model setup is three-layered. Grok four point seven does the implementation, then GPT six point one and Opus five point five review it. Merging happens on its own only when all three say the code is safe.

James: And he's candid, Maya. Paying for all three subscriptions is not cheap, though he calls it cheaper than hiring an engineer. He also says the setup is not perfect and makes some mistakes. Honest caveats for something merging his code while he sleeps.

Maya: Next, James, Cursor has no built-in Unity connector, so a walkthrough shows setting it up by hand. When you create a new project, check Use Unity CLI, install Node, then paste the install command from the official Unity skill repo and confirm.

James: And the clever part, Maya, is the conflict fix. Cursor tried to use MCP because it had read the MCP skill. The fix is npx skills remove dash g, find the MCP skill, remove it. Then Cursor read the Unity CLI skill and finished the scene.

Maya: Exactly, and you choose global or project-only install, with symlink as the method. The walkthrough also suggests an agents dot md telling the agent to always use the Unity CLI, never touch meta files, and follow a stated project structure.

James: That agents dot md detail is the glue, Maya. Without it the agent drifts. With it, you encode the rules once and every session reads them before it touches your Unity project.

Maya: Now a sharp find on r slash cursor, James. One user argues Cloud Agents already run on a machine your plan pays for, so a spare always-on box is unnecessary. The catch is where your setup lives, since the agent boots a project clone.

James: And here's the split, Maya. Skills placed in home dot cursor skills worked first try. Slash commands in home did not. They only worked from the repo's dot cursor commands folder. The user calls it a path problem, not a hardware problem.

Maya: Right, and his entry point is a committed dot cursor environment dot json plus a pinned image. He's clear this is one report from a single setup, and he asks whether others see the same skills-versus-commands split. A caveat worth keeping, James.

James: Last one, Maya. A post from October first found references to grok dash four point eight inside Cursor's internal list of selectable models, and in a test file in xAI's shared Grok Build repository. That suggests engineers are wiring it in before any public release.

Maya: And the timing is notable, James, because Grok four point seven had a bumpy rollout. Musk said September second it was about ten days out, then September eleventh a few more days. It shipped on x dot ai September twenty-first, now live in Cursor.

James: Fair, and the post is careful, Maya. No official model page, price, or release date exists for four point eight. The takeaway is to watch for an official grok dash four point eight model ID in the picker, not treat pre-release references as a release date.

Maya: That's your Cursor HQ daily brief. Check cursorhq.ai for the full stories. Keep building.