June 21, 2026
GLM 5.2, Codex Skills, Claude & Cursor
GLM 5.2 passed the vibe check, Codex Record and Replay makes screen recordings into skills, Cursor got much more interesting after the SpaceX acquisition, and Claude pushed design mode plus mini apps forward.
This episode is about the week AI agents got pulled in four directions at once: cheaper open models, better skill creation, Cursor getting a serious platform boost, and Claude pushing more of the app layer forward.
The most practical idea is simple: if you want to become Agent Native, you should be testing the tools that make agents cheaper, easier to train, and easier to use inside real workflows.
The week in one view
| Update | Why it matters |
|---|---|
| GLM 5.2 | A cheaper open model that feels close enough to frontier models to test in real workflows. |
| Codex Record and Replay | Lets you record a workflow and turn it into a reusable agent skill. |
| SpaceX acquiring Cursor | Gives Cursor the resources to compete as a broader AI agent platform. |
| Claude design and artifacts | Moves Claude toward visual editing, sharing, and agent-native mini apps. |
GLM 5.2 actually passed the vibe check
Z.AI released GLM 5.2, an open model that is much cheaper than frontier models and still feels close enough to Opus 4.8 and GPT 5.5 to take seriously.
The important part is that this was not just a benchmark hype cycle. After testing it inside Cursor and with real agent tasks, it felt good enough to use for actual work.
Open models change the agent cost story
If open models keep getting this good, the cost of running agents all day starts to look very different. That matters for personal agents, business automations, and any workflow where the model has to run for a long time.
The bigger signal is that the frontier labs now have real pressure from open source. That competition should push OpenAI, Anthropic, Google, and the open model labs to move faster.
Codex Record and Replay turns demos into skills
Codex's Record and Replay workflow is underrated. You can record yourself doing a task, let Codex inspect the recording, and turn that process into a reusable skill.
That is a big deal because most useful agent work is not just one prompt. It is teaching the agent the messy steps you already know how to do across apps like Typefully, Comet, files, and your browser.
Cursor is becoming a real super app contender
The SpaceX acquisition matters because Cursor now has the capital, compute, distribution, and model-training resources to compete harder with Codex and Claude.
The best outcome for users is a tight race between Codex, Claude, Cursor, and open source models. More competition means better products and better pricing for people actually using agents.
Claude is pushing the app layer
Claude's design mode update makes it easier to edit directly on the canvas, stay closer to a design system, and send work into other tools.
Claude Code artifacts are the bigger signal. They turn agent output into shareable mini apps, which is exactly the direction agent-native software is going.