Zhipu's ZCode Agent Gives GLM5.2 a Frontend Boost That Claude Code Couldn't Match
The same model can produce radically different output quality depending on the agent runtime that wraps it. ZCode's results suggest that tooling, prompt scaffolding, and execution environment may matter as much as the model itself for frontend generation tasks, a signal worth watching as Chinese AI labs ship their own Codex competitors.
ZCode is Zhipu's desktop agent coding tool, now at version 3.0.0, with a UI that closely resembles OpenAI's Codex. It supports Windows and macOS, offers three login paths (international z.ai, domestic BigModel, and API key), and currently gives new users five free days with subscribers getting a 150% quota bump. The tool picks a project folder, sets model and thinking budget, then generates code through a chat-driven agent loop.
In side-by-side tests against Claude Code using the same GLM5.2 model, ZCode produced working output where Claude Code repeatedly hung or threw API errors. A complex text-adventure prompt requiring ASCII art, butterfly-effect state tracking, and humor-based input rejection ran for 17 minutes in ZCode and delivered a polished result with loading animations. A second test, a cyberpunk reimagining of a classical Chinese painting, generated 506 lines of changes in about 10 minutes with layout and color work rated well above average.
The gap is most visible in frontend output. The same model inside ZCode produced UI and interaction quality that the tester, who had previously found GLM's frontend work weak, called "a completely different model." Whether this is a systematic advantage of ZCode's agent scaffolding or just lucky sampling remains an open question the tester plans to answer with a full rerun of all benchmarks.
The wide output-quality gap between the same model in different agent shells implies that scaffolding, prompt structure, and execution environment are first-class variables in AI coding performance, not just the model card.
ZCode's strong frontend results challenge the assumption that GLM's models are inherently weak at UI generation; the bottleneck may have been the runtime, not the model weights.
A 17-minute completion time for a complex prompt is long enough to be impractical for rapid iteration but still a working result where the alternative produced nothing, resetting expectations for patience in agent workflows.
The discussion centers on ZCode's severe capacity problems and a sharp split over its real-world utility. Several developers report being unable to subscribe to plans or hitting constant rate limits and congestion, with one attributing it to genuine compute shortages rather than deliberate throttling. A detailed critique contrasts ZCode's strong performance on large feature requests with its disastrous looping behavior on small bug fixes, where it burned far more time and tokens than Trae on the same GLM5.2 model. A separate thread questions whether the article was plagiarized, which the author denies.
The ZCode workflow feels overly complicated. For a big feature request the results are actually pretty good, but asking it to fix small bugs is an absolute disaster. It always gets stuck in loops. It even says in its thinking process, 'I'm overcomplicating this, I shouldn't loop,' and then keeps looping anyway. Both connected to 5.2, same prompt, just fixing a focus-loss issue. Trae solved it in 3 minutes; ZCode looped for half an hour. I tried three times in disbelief and got the same result. In real-world use, even with 1.5x the token allowance, it's just not as durable as other tools.
This company's model has no restrictions for overseas users but deliberately does hunger marketing targeting Chinese users! Thumbs down! It keeps reporting congestion while you're using it, deliberately reconnecting and retrying!
It really is insufficient computing power. Codingplan Max members often get rate-limited and slowed down because too many people are using it.
I'm here to complain. Can't even subscribe to TokenPlan.