Zhipu ZCode Upgrade: Four New Features Push China's Coding Harness Into Autonomous Delivery Era
> Sourced via AI HOT (WeChat: 智谱 GLM, 2026-08-11) > Secondary sources: Beijing Daily, Weibo @智谱, mtrt.cn, dtm.com.cn, ITBear
Key Points
1. Boundaries of the four new features
Zhipu shipped four upgrades to ZCode on August 11, 2026, while announcing the product had surpassed one million users. This marks the first time a domestic Chinese Coding Harness has positioned autonomous delivery as a shipping product surface rather than a slogan.
| Feature | Problem it solves | Key constraints | |---|---|---| | Goal mode | Agent runs complex tasks end-to-end instead of waiting on human pushes at each step | Goals must be explicit and verifiable (e.g., "first paint ≤ 2s and all tests pass") | | Subagents | Long tasks pollute the main conversation context | Built-in General-purpose + Explore types; users can create custom ones | | Remote Control | Users want to walk away from the desk during long tasks | Phone is only a controller; code execution stays on the desktop | | Idle-Time Tasks | Cost of slow but non-urgent jobs | Runs in off-peak windows; does not consume GLM Coding Plan credits |
Goal is the core change. It moves developers from the foreman seat to the client seat — hand the agent an acceptance criterion, and the agent handles task decomposition, code edits, command execution, test inspection, and retry loops on its own.
2. Performance numbers: the +2.39% pass-rate gain is not noise
Zhipu published hard benchmark data on its in-house Z.ai Code Bench:
- GLM-5.2 + ZCode vs GLM + Claude Code: overall task pass rate higher by 2.39%
- Stronger lead on complex tasks: cross-file, cross-stage, acceptance-driven jobs are the main battleground
- Checklist-item pass rate lower by 1.22%: Claude Code wins on narrow, single-point checks, but GLM is more stable on full closure
- Cache hit rate above 98%: drives roughly 30% more effective tokens
- For individual developers: effective credits expand by roughly 1.5–1.8x
- For enterprise teams: overnight idle compute becomes real throughput without crowding production-window quotas
Combined with a 1.5x limited-time quota boost running August 11–31, real-world usable resources approach 1.8x the normal level.
The implication: the same base model wrapped in different harnesses can produce materially different outcomes. Zhipu's claim — "the model sets the capability ceiling; the harness handles context management, tool calls, task scheduling, caching, and result verification, and decides how much of that ceiling is actually realized" — is now backed by its own benchmark.
3. Crossing one million users is a watershed
"ZCode surpasses one million users" matters more than the feature list itself.
Domestic Chinese coding tools have historically been stuck at the "can it run a single task" threshold. Running ≠ being usable; being usable ≠ being adopted. One million users signals:
1. The subscription model works: GLM Coding Plan is subscription-based, not usage-metered, putting it in the same business-model tier as Cursor, Claude Code, and Codex. 2. The ecosystem flywheel is spinning: more users → more data → better harness scheduling policies. 3. The contest shifted from base models to toolchains: DeepSeek, Qwen, and GLM all have strong bases; differentiation now lives in the harness.
The real signal to the industry is not the feature list but the formal assembly of the base model + Coding Harness + subscription service three-piece ecosystem.
4. The hidden value of Subagents
Subagents do more than parallelize work — they address a problem Harness Engineering has underweighted: main-conversation context pollution from long tasks.
Under Goal mode, "automatic decomposition + cross-file refactor" runs that last dozens of minutes are normal. If the main conversation keeps absorbing that traffic, token consumption grows linearly and attention gets dragged into old context that is no longer relevant. Subagents let Explore-type read-only children handle discovery, General-purpose children handle edits, and the main conversation hold only the goal and termination conditions.
This raises Harness design complexity again: *what context is shared, what is isolated, when is it reclaimed* between parent and child agents is a new engineering category. Claude Code uses worktree isolation; Codex uses session-ID addressing. Zhipu's path is a middle route — main conversation + two built-in types + user-created custom agents. Whether this lands real data during the 1.5x boost window by month-end is the open question.
5. Idle-Time Tasks: a differentiator in cost structure
Idle-Time Tasks looks small but is actually a commercial differentiator.
GLM Coding Plan is credit-based — once credits run out, the session stops. Idle-Time Tasks run in low-traffic server windows without consuming credits, so for long, non-urgent jobs (code review, performance optimization, large-scale testing), real cost approaches zero. This opens a gap in the all-tasks-pay-per-use cost structure used by Cursor and Claude Code.
6. The real gap versus Cursor and Claude Code
Three layers of gap remain:
1. Model ecosystem breadth: Claude Code runs the full Claude family (Haiku / Sonnet / Opus 4.6); ZCode runs GLM. Model diversity and global deployment capacity differ. 2. International ecosystem integration: Claude Code and Codex have native hooks into GitHub Issues, PR Review, and Slack; ZCode needs time to reach parity. 3. Enterprise compliance: finance and government clients need private deployment and data compliance — a category where domestic Chinese harnesses are actually advantaged.
Conversely, ZCode's leads in localization, China-specific deployment, and subscription pricing are real. If the one-million-user cohort retains at 30%+, the domestic AI coding landscape becomes not "one or two winners" but a four-way split among Zhipu, Alibaba Qwen, DeepSeek, and Kimi.
---
Core data: 1M users, 4 new features, +2.39% task pass rate, 98%+ cache hit rate, +30% effective tokens, 1.5x limited-time boost → ~1.8x effective capacity, −1.22% checklist-item pass rate Timeline: 2025 project kickoff → 2026.08.11 upgrade + user milestone → 2026.08.31 boost window ends Sources: 智谱 GLM WeChat account (2026-08-11), Beijing Daily, ITBear, mtrt.cn