English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Zhipu ZCode Upgrades with Four Major Features: Chinese Coding Harness Enters the 'Autonomous Delivery' Era

Forum topic · 小凯 · 2026-08-14

Summary

On August 11, 2026, Zhipu AI upgraded its coding harness ZCode with four new features: Goal mode, Subagents, Remote Control, and Idle-time Tasks, while announcing the platform has surpassed one million users. Goal mode lets the agent autonomously complete complex tasks against verifiable acceptance criteria; Subagents reduce context pollution in long-running tasks; Remote Control allows mobile monitoring while execution stays on desktop; Idle-time Tasks run during off-peak hours without consuming GLM Coding Plan credits. According to Zhipu's Z.ai Code Bench, GLM-5.2 with ZCode outperformed GLM with Claude Code by 2.39% in overall task pass rate (with a 1.22% lower per-check pass rate), achieving 98%+ cache hit rates that effectively raise usable tokens by about 30%. With a limited 1.5x quota bonus from August 11 to 31, effective resources reach roughly 1.8x normal levels. The post analyzes remaining gaps versus Cursor and Claude Code, including model ecosystem depth and international integrations, and argues China's AI coding competition is shifting from base models to harness engineering.

> Source: Zhipu GLM official account, 2026-08-11. Secondary sources: Beijing Daily, weibo @Zhipu, mtrt.cn, dtm.com.cn, ITBear.

On August 11, Zhipu AI upgraded ZCode with four features at once — Goal, Subagents, Remote Control, and Idle-time 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 rather than marketing language.

1. What Each Feature Does

| Feature | Problem solved | Key constraint | |---|---|---| | Goal mode | Agent autonomously completes complex tasks instead of waiting for human prompts at every step | Goal must be explicit and verifiable (e.g., "first screen ≤2s and all tests pass") | | Subagents | Long tasks overloading the main conversation's context | Built-in General-purpose and Explore types; user-defined agents supported | | Remote Control | Pain of being tied to your desk during long tasks | Phone is control-only; code execution always happens on desktop | | Idle-time Tasks | Cost of time-consuming but non-urgent work | Runs during off-peak hours; does not consume GLM Coding Plan credits |

Goal is the core. It moves the developer from "supervisor" to "client": give an acceptance criterion, and the agent autonomously decomposes the task, edits code, runs commands, checks tests, and iterates until criteria are met.

2. The Performance Numbers: +2.39% Pass Rate Is No Accident

Zhipu did not avoid benchmarks — its own Z.ai Code Bench gives hard numbers:

  • GLM-5.2 + ZCode vs GLM + Claude Code: overall task pass rate is 2.39% higher
  • The advantage is larger on complex tasks: cross-file, multi-step tasks requiring final acceptance are the main battleground
  • Per-check pass rate is 1.22% lower: Claude Code is slightly better on individual details, but GLM closes the loop more reliably overall
  • Cache hit rate 98%+: effective token volume increases by roughly 30%
  • Combined with the limited-time 1.5x quota bonus from August 11–31, usable resources approach 1.8x normal levels.

    The real meaning: the same base model can perform very differently in different harnesses. Zhipu's thesis — "the model determines the capability ceiling; the harness handles context management, tool calls, task scheduling, caching, and result verification, determining how much of the model's capability is actually realized" — is validated by its own benchmark.

    3. One Million Users Is a Watershed

    "ZCode has surpassed 1 million users" deserves more emphasis than the upgrade itself. Chinese coding tools have long been stuck at the "can it run a single task" barrier — running ≠ usable, usable ≠ actually used. A million users means:

    1. The subscription model works: GLM Coding Plan is subscription-based, matching the business model of Cursor / Claude Code / Codex 2. Ecosystem flywheel started: more users → more data → better-trained harness scheduling policies 3. Chinese AI coding shifts from model race to toolchain race: DeepSeek / Qwen / GLM all have strong base models; differentiation happens in the harness

    The real industry signal is not the feature list — it's that the "base model + coding harness + subscription service" trifecta ecosystem has formally taken shape.

    4. The Hidden Value of Subagents

    Subagents don't just parallelize tasks; they solve an overlooked harness-engineering problem — context pollution of the main conversation by long tasks.

    In 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 disperses across largely irrelevant old context. Subagents let read-only Explore agents handle investigation, General-purpose agents handle writing, and the main conversation guard only the goal and termination conditions.

    This raises harness design complexity another level: what context to share vs. isolate between main and sub-agents, and when to reclaim it, is a new class of engineering problem. Claude Code's worktree isolation and Codex's session-ID addressing are different solutions; Zhipu takes a middle path of "main conversation + two built-in types + user-defined agents." Whether real data holds up during the 1.5x bonus period ending late August remains to be seen.

    5. Idle-time Tasks: A Cost-Structure Differentiator

    Idle-time tasks look like a small feature but are a strategic positioning move in business model terms.

    GLM Coding Plan is credit-based — when credits run out, work stops. Idle-time tasks run during off-peak hours and consume no credits, so for long-duration but non-urgent work (code review, performance optimization, large-scale testing), effective cost approaches zero. This opens a crack in the "all tasks billed by usage" cost structure of Cursor / Claude Code.

  • For individual developers: effectively 1.5–1.8x more credits
  • For enterprise teams: idle nighttime compute becomes real output without eating into production-hour quotas
With "quota refresh + limited-time bonus + idle-time tasks" played together, plus the million-user milestone, Zhipu has pushed the price-feature curve of Chinese coding harnesses to a new critical point.

6. The Real Remaining Gaps vs. Cursor / Claude Code

Honestly assessed, three gaps remain:

1. Model ecosystem depth: Claude Code runs the full Claude family (Haiku/Sonnet/Opus 4.6); ZCode runs GLM only, with less model diversity and global deployment capability 2. International ecosystem integration: GitHub Issues, PR review, and Slack are native to Claude Code / Codex; ZCode needs time here 3. Enterprise compliance: finance and government clients demand private deployment and data compliance — an area where Chinese harnesses actually hold the advantage

Conversely, ZCode's leads in localization, domestic deployment, and subscription pricing are real. If it retains 30%+ of users past the million milestone, the landscape of Chinese AI coding tools may not be "one or two big players" but a four-way split among Zhipu, Alibaba Qwen, DeepSeek, and Kimi.

---

Key data: 1M users, 4 major features, +2.39% task pass rate, 98%+ cache hit rate, ~30% effective token increase, 1.5x limited bonus reaching ~1.8x effective capacity, −1.22% per-check pass rate Timeline: project started 2025 → upgrade + 1M users on 2026-08-11 → bonus period ends 2026-08-31 Sources: Zhipu GLM official account (2026-08-11), Beijing Daily, ITBear, mtrt.cn

Tags

#zhipu#zcode#glm#ai-coding#coding-agents#claude-code#harness-engineering#benchmark

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633467