> Originally published via Zhipu GLM official account (2026-08-11). Secondary sources: Beijing Daily, Weibo @Zhipu_AI, mtrt.cn, dtm.com.cn, ITBear.
1. The Boundaries of the Four Features
On August 11, Zhipu upgraded ZCode with four features in one release—Goal, Subagents, Remote Control, and Idle Tasks—while pushing its user count past one million. This is the first time a Chinese coding harness has delivered "autonomous delivery" as an actual product form, not just marketing language.
| Feature | Problem solved | Key constraint | |---|---|---| | Goal mode | Agent completes complex tasks autonomously instead of waiting for manual push at every step | Goals must be explicit and verifiable (e.g., "first paint ≤2s and all tests passing") | | Subagents | Long tasks over-consuming the main conversation's context | Built-in General-purpose and Explore types; users can create their own | | Remote Control | Pain of being tied to the desk during long tasks | Phone is control-only; code execution always stays on desktop | | Idle Tasks | Cost of time-consuming but non-urgent work | Runs during off-peak hours; consumes no GLM Coding Plan credits |
Goal mode is the core. It moves developers from "supervisor" to "client"—give one acceptance standard, and the agent autonomously decomposes, edits code, runs commands, checks tests, and iterates until the bar is met.
2. The Performance Numbers: +2.39% Pass Rate Is No Accident
Zhipu didn't shy away from benchmarking; its in-house Z.ai Code Bench produced hard numbers:
- GLM-5.2 + ZCode vs GLM + Claude Code: overall task pass rate 2.39% higher
- Advantage widens on complex tasks: cross-file, multi-stage tasks requiring final acceptance are the main battlefield
- Check-item pass rate is 1.22% lower: Claude Code slightly wins on single-point details, but GLM closes the loop more reliably overall
- Cache hit rate 98%+: effective token volume increases by roughly 30%
- For individual developers: effective credits expand by 1.5–1.8×
- For enterprise teams: overnight idle compute becomes real output without eating into production-window quotas
Combined with the limited-time 1.5× credit boost from August 11–31, users' effective resources approach 1.8× normal levels.
The real takeaway: the same base model can perform very differently in different harnesses. Zhipu's thesis—"the model sets the capability ceiling; the harness handles context management, tool calling, task scheduling, caching, and result verification, deciding how much of that capability is actually realized"—is validated by its own benchmark.
3. One Million Users Is a Watershed
"ZCode surpasses 1 million users" deserves more emphasis than the upgrade itself. Chinese coding tools have long been stuck at "can it complete a single task"—running ≠ usable, usable ≠ actually used. A million users means:
1. The subscription model works: GLM Coding Plan is subscription-based, on par with Cursor / Claude Code / Codex business models 2. A positive ecosystem loop has started: more users means more data, which sharpens harness scheduling policies 3. Chinese AI coding is shifting from model competition to toolchain competition: DeepSeek / Qwen / GLM all have strong base models; the differentiation battlefield is the harness
The real signal to the industry isn't the feature list—it's that the "base model + coding harness + subscription service" trio ecosystem is officially taking shape.
4. The Hidden Value of Subagents
Subagents aren't just about parallelism; they address an overlooked harness-engineering problem—long tasks polluting the main conversation's context.
Under Goal mode, "auto task decomposition + cross-file refactoring" routinely runs for tens of minutes. If the main conversation carries all of it, token consumption grows linearly and attention dilutes across stale context. Subagents let read-only Explore agents handle investigation, General-purpose agents handle writing, and the main conversation only guard the goal and termination conditions.
This pushes harness complexity up another level: what to share, isolate, and reclaim between main and sub-agents is a new class of engineering problem. Claude Code's worktree isolation and Codex's session-ID addressing are different solutions. Zhipu's middle path—"main conversation + two built-in types + user-defined agents"—will be tested by real data once the 1.5× boost period ends in late August.
5. Idle Tasks: A Cost-Structure Differentiator
Idle Tasks look minor but represent a business-model position play.
GLM Coding Plan is credit-based—when credits run out, work stops. Idle Tasks run in off-peak windows and consume no credits, meaning long-duration but non-urgent work (code review, performance tuning, large-scale testing) costs users near zero. That opens a crack in the pay-per-use cost structure of Cursor / Claude Code.
6. The Real Gap vs. Cursor / Claude Code
Being objective, three gaps remain:
1. Model ecosystem depth: Claude Code runs the full Claude family (Haiku/Sonnet/Opus 4.6); ZCode runs GLM only—model diversity and global deployment lag 2. International ecosystem integration: GitHub Issues, PR review, and Slack are native for Claude Code / Codex; ZCode needs time here 3. Enterprise compliance: finance and government customers demand private deployment and data compliance—where Chinese harnesses actually hold an advantage
Conversely, ZCode's lead in localization, domestic deployment, and subscription pricing is real. If it can retain 30%+ of users past the million-user milestone, the Chinese AI coding landscape won't be "one or two winners" but a four-way split among Zhipu, Alibaba Qwen, DeepSeek, and Kimi.
---
Key data: 1M users, 4 features, +2.39% task pass rate, 98%+ cache hit rate, ~30% effective token gain, 1.5× boost ≈ 1.8× effective capacity, -1.22% check-item pass rate Timeline: 2025 project start → 2026-08-11 upgrade + 1M users → 2026-08-31 boost period ends Sources: Zhipu GLM official account (2026-08-11), Beijing Daily, ITBear, mtrt.cn