On August 11, Zhipu AI announced a major upgrade to ZCode, its self-developed coding harness, officially launching four new features: Goal, Subagents, Remote Control, and idle-time tasks. As a thank-you for passing the one-million-user milestone, all GLM Coding Plan users had their quotas reset to full at 13:00 that day. Original link: https://zcode.z.ai/cn
It's Not the Same Species as Claude Code
ZCode is Zhipu's homegrown Chinese-developed coding harness — not an IDE wrapper, not a ChatGPT wrapper, but a framework deeply optimized for GLM models covering "context management + tool calling + task scheduling + caching + result validation." Zhipu internally defines the harness as the middle layer that "determines how much of the model's capability can actually be realized" — the same GLM-5.2 running in ZCode versus Claude Code shows markedly different real-world results.
Zhipu built Z.ai Code Bench to quantify this. Based on real user scenarios across three subtask categories (Full Stack, Bug Fix, Feature Implementation), it constructs simulated local programming environments and evaluates along four dimensions: regression tests, new-feature tests, front-end simulated interaction, and code quality assessment. Its stated design goal is to "avoid the data contamination that public leaderboards may suffer" — a direct response to SWE-bench's evaluation gaming.
The Real Gap: 2.39%
On Z.ai Code Bench, GLM-5.2 + ZCode beats GLM-5.2 + Claude Code by 2.39% on overall task pass rate — but is 1.22% lower on checklist pass rate.
The combination is interesting: ZCode is stronger at "finishing the task end-to-end," while Claude Code is stricter at "matching every checklist item." Zhipu's own interpretation: ZCode's advantage is concentrated in complex tasks requiring cross-file, multi-stage collaboration and final acceptance. For real production scenarios (a PR touching 5 files + CI + unit tests), ZCode's end-to-end pass rate is higher, but it's looser on per-step format strictness (log formats, naming conventions, etc.).
The Economics of a 98% Cache Hit Rate
ZCode made context-cache reuse a dedicated optimization, achieving a cache hit rate above 98% for GLM (official figure: 98.10%). This boosts the effective token volume of the GLM Coding Plan by roughly 30% — repeated context is billed at a lower credit coefficient, so the same money runs more tasks.
Stacked with the limited-time 1.5x quota bonus running through August 31, overall usage approaches 1.8x the regular quota. This is Zhipu's price signal in the coding-subscription race — running GLM in ZCode before end of August is nearly a 50% price cut.
Goal Mode: From Babysitting Agents to Setting Goals and Waiting
This is the core of the upgrade. The old coding-agent workflow was:
> Developer sends request → waits for agent → pushes it forward → tests fail → follow-up questions → discover omissions → supply more context → push again
Goal mode changes it to:
> Set "first-screen load ≤ 2s" + "all existing tests pass" → ZCode decomposes automatically → edits code → runs commands → executes tests → checks results → iterates if unmet → until done
Each round's progress, duration, and results are visualized in the Goal panel. It's essentially bringing the RL paradigm of "sparse reward + many trials" into a coding-agent product. Code review is pushed to after goal completion.
Subagents: A Parallel Dev Team
Under Goal mode, Subagents turn one agent into a parallel development team. Two built-in types:
- General-purpose: modify code, fix issues, run commands
- Explore: read-only exploration — locate code, analyze call chains, gather pre-change evidence
- Entry points: WeChat, Feishu, Lark — via QR code or browser link
- Phone can: view progress, input commands, create new tasks, reconnect a dropped workspace
- Phone cannot: code and commands are never uploaded — they still run in the desktop's original environment (local / SSH / WSL / Docker)
- Key point: no cloud environment is created; projects are not synced to the phone
- Overall task pass rate, GLM-5.2 + ZCode vs + Claude Code: +2.39%
- Checklist pass rate: -1.22%
- Cache hit rate: 98.10%
- Effective token volume increase: ~30%
- Limited-time quota bonus: 1.5x (through 2026-08-31)
- Combined effective usage: about 1.8x regular quota
- ZCode users: 1 million+ (announced 2026-08-11)
- Built-in subagent types: General-purpose, Explore
- Remote Control entry: WeChat, Feishu, Lark, browser link
- Workspace compatibility: local / SSH / WSL / Docker
- Foundation model: GLM-5.2 (Zhipu in-house)
- Harness: ZCode (Zhipu in-house, optimized for GLM)
- Subscription: GLM Coding Plan (Zhipu-operated)
- https://zcode.z.ai/cn
- https://mp.weixin.qq.com/s?__biz=MzkyMzI3NzQ0Mg%3D%3D&mid=2247494052&idx=1&sn=ee3ab3d0f4550e9120927c53a27522c9 (Zhipu official WeChat account)
- https://finance.sina.com.cn/tech/digi/2026-08-11/doc-inimxmep7594993.shtml (Sina Finance, via ITHome)
- https://www.chinaz.com/ainews/30242.shtml (Chinaz)
- https://www.163.com/dy/article/L42IHT8M053469RG.html (NetEase Tech)
Users can also define custom subagents with configurable model, permission scope, and prompts, invoked directly in the chat box via "/". For cross-file refactors, complex bug hunts, and full-stack features, "one agent does everything" is replaced by "multiple agents divide the work" — reducing main-context load on long tasks.
Remote Control: Phone Control Without Uploading Code
Remote Control temporarily exposes the current ZCode desktop window to a phone:
It's the "remote work" version of a coding agent — push your agent forward away from your desk while code never leaves your machine.
Idle Tasks: Long Jobs at Low Peak, No Credits Deducted
Idle tasks let users submit non-urgent, long-running jobs (large test suites, code optimization). ZCode executes them automatically during low-peak hours, and within entitlement limits does not deduct GLM Coding Plan credits. It's cloud computing's "off-peak discount" brought to the agent layer — for developers, the marginal cost of long tasks approaches zero.
Key ZCode Numbers
Zhipu's "Foundation Model + Harness + Subscription" Trinity
ZCode is not another Claude Code clone. It closes a loop of Chinese-developed foundation model + harness + subscription service:
Zhipu has been steadily filling out its foundation — models, agents, MaaS, industry ecosystem — and ZCode is the entry point closest to developers: a million-user base plus cache economics plus Goal/Subagent autonomous delivery.
Compare with Claude Code: Anthropic's harness, Anthropic's model, Anthropic's subscription — all three coupled in one company. ZCode claims all three layers for itself, but this path must face Claude Code's moat in Western developers' minds.
The signal most worth watching: will Zhipu make Z.ai Code Bench public (currently an internal benchmark) so third parties can independently verify the 2.39% gap? If it's published and independently reproduced, it becomes a turning point for Chinese AI coding tools — if it stays internal, the "2.39% ahead of Claude Code" claim should be taken with a discount.
---
Sources