On August 11, 2026, Zhipu (02513.HK) announced a major ZCode upgrade -- ZCode 3.0 -- with four new features going live: Goal (goal mode), Subagents, Remote Control, and idle-time tasks. At 13:00 the same day, as a milestone reward for crossing one million users, GLM Coding Plan quotas were reset across the board for all users, fully refilled. This is not an ordinary feature iteration. The Claude Code backdoor incident is still reverberating; one team migrated everyone to ZCode 3.0 within 48 hours, and foreign media are already asking whether Chinese model companies are making a grab for the AI coding entry point. ZCode 3.0 uses Goals plus Subagents to push the coding agent from "line-by-line completion" toward "goal-level autonomous delivery," then layers on a quota refill for acquisition. The timing and product cadence are well-placed. Set against the three-camp landscape in our AI coding agent landscape overview, this is the domestic camp's key move to seize the entry point.
A boundary note first: this article is assembled from Sina Finance, Caiwen, Yunbaba, ZAKER, Zhihu, Tencent Cloud, and other public reports, current as of 2026-08-11. ZCode 3.0's exact capability boundaries, quota values, and rules are subject to Zhipu's official documentation. This is trend analysis, not internal material, and no independent hands-on reproduction was performed.
1. Four New Features in ZCode 3.0: From Completion to Autonomous Delivery
The core of this upgrade is moving the developer's intervention point from "every line of code" up to "every goal." Each of the four features solves a specific bottleneck.
Goal (goal mode) tackles instruction granularity. Instead of describing changes line by line, the developer gives a goal-level description, such as "add a resumable upload endpoint." The agent autonomously handles requirement decomposition, file location, code modification, and self-test verification. The intervention point shifts from "reviewing every diff line" to "accepting every goal." This is the critical leap from copilot to agent: a copilot helps you write code, an agent delivers features for you.
Subagents tackles the context-window ceiling of a single agent. Complex tasks are split across multiple collaborating agents, each owning its lane, with a main agent coordinating. No matter how large a single agent's context window, it has a limit; splitting lets each sub-agent focus on its own domain, keeping context cleaner and decisions sharper.
Remote Control (bot remote control) tackles the lock-in of sitting in your IDE to monitor progress. The agent plugs into Feishu, WeChat, and Telegram, so you check progress, issue commands, and trigger sub-tasks from IM. This opens an asynchronous collaboration channel for long-running tasks: the agent works while you ask "where is it at?" in a group chat and it answers.
Idle-time tasks tackles the session-break problem in long-running work. When the developer is free, the agent pushes tasks forward in the background, with state, context, and decisions all persisted, surviving across sessions and across days. Meetings during the day, a refactor running overnight, results waiting in the morning -- that is the real development rhythm.
2. Feature Comparison: What Each One Fixes
| Feature | Pain point solved | Intervention shift | Practical meaning |
|---|---|---|---|
| Goal mode | Line-by-line instructions too granular | Every line -> every goal | From completion to autonomous delivery |
| Subagents | Single-agent context overflow | Single-thread -> multi-agent division | Complex tasks split by domain, parallel |
| Remote Control | Must sit in IDE to watch progress | Sync monitoring -> async control | Check progress, send commands from IM |
| Idle-time tasks | Long tasks break across sessions | One-shot -> persisted | Push across days, collect results later |
On the technical side, ZCode 3.0 runs on Zhipu's self-developed agent kernel, deeply adapted to GLM-5.2. The broader capability set actually spans seven modules: Goals, Bot remote control, long-running tasks, multi-end collaboration, full-context awareness, automated code review, and MCP with Skill extensions. The four features highlighted this time are the ones from the seven modules that best demonstrate the "autonomous delivery" difference.
3. Timing: The Claude Code Backdoor Window
Beyond the features, timing is the real story of this article. The Claude Code backdoor incident pushed some developer teams to reassess their dependence on a single overseas coding tool. One team migrated everyone to ZCode 3.0 within 48 hours -- that speed signals a stress-driven response to a security event, not a deliberate long-term selection.
In this window, domestic coding tools need two things to grab attention: the product must genuinely keep up, and migration cost must be low. ZCode 3.0 launching Goals and Subagents at this moment is saying "I am not just a substitute; I have taken a step further on the autonomous delivery path." Foreign media attention -- "Chinese model companies are starting to grab the AI coding entry point" -- shows this is being framed inside a larger competitive narrative. ZCode 3.0 has been called a "domestic Codex alternative" by some voices; that label came from the market, not from self-declaration.
But objectively, the precondition for seizing the entry point is a product that can deliver. Whether Goal mode's autonomous decomposition is accurate, whether Subagents collaboration is stable, whether idle-time task persistence loses context across days -- these must be validated on real projects. The window gives you traffic; whether you keep it depends on the product.
4. The Quota Refill: An Overt Acquisition Play
GLM Coding Plan quotas for all users were reset at 13:00 on 8/11, fully refilled, under the banner of a "one-million-user milestone reward." ZCode crossing one million users is a milestone, but the milestone itself is not the point; the point is using the quota refill to bring old users back and get new users to try.
Quota is the ammunition of a coding tool. After the Claude Code backdoor incident, migrating teams' biggest fear is "we migrate and then find the quota is not enough to run long-running tasks." ZCode 3.0 is pushing idle-time tasks and long-running tasks right now -- both are heavy quota consumers. The quota refill and the feature upgrade announced on the same day is no coincidence: you need ammunition to run the multi-step tasks that Goal decomposes, to sustain the cross-day background consumption of idle-time tasks. Features give you a reason to use it; quota gives you the capacity to use it. It is a combined punch in product cadence. Specific quota values and refill rules are subject to Zhipu's official documentation.
5. Practical Impact for Developers
If you are already on the GLM stack, ZCode 3.0's migration cost is the lowest, and Goals and Subagents can be trialed directly on existing projects. Three things to remember when starting: validate Goal mode's decomposition quality with small goals first, do not throw a big requirement at it on day one; draw clear task boundaries for Subagents, or multiple agents trampling each other's context is worse than a single agent; and lock down the IM command permissions for Remote Control, so the agent is not accidentally triggered to run wild in a group chat.
If you are a team migrating from Claude Code, evaluate two things: first, ZCode 3.0's adaptation to your existing stack (languages, frameworks, build tools), since Goal mode's autonomous decomposition depends heavily on project understanding depth; second, GLM-5.2's code generation quality in your specific business scenarios. Model capability is the foundation, features are the superstructure; if the foundation is weak, flashy features are useless.
For anyone doing tech selection, the signal from this upgrade is that domestic AI coding tools have moved past the "good enough as a substitute" stage and are pushing a differentiated "autonomous delivery" route. But do not get swept up by feature lists. Three hard metrics decide whether it is production-ready: long-running task context-persistence reliability, multi-agent collaboration stability, and Goal decomposition accuracy. All three must be validated on your real projects; do not trust demos.
6. FAQ
Q: How does ZCode 3.0's Goal mode differ from ordinary code completion? A: Completion means you write a line and it continues, with the intervention point at every line. Goal mode means you give a goal-level description (such as "add a resumable upload endpoint"), and the agent autonomously handles requirement decomposition, file location, code modification, and self-test verification, moving the intervention point up to every goal. It is the shift from helping you write code to delivering features for you.
Q: Is the GLM Coding Plan quota refill available to everyone? A: Per public reports, quotas were reset across the board at 13:00 on 8/11, fully refilled, as a one-million-user milestone reward. Specific quota values and rules are subject to Zhipu's official documentation.
Q: Is Remote Control safe to connect to IM? A: Functionally it supports Feishu, WeChat, and Telegram for checking progress and sending commands. On the safety side, lock down IM command permissions to avoid accidental triggers. The specific permission model is subject to official documentation; do not connect it raw to a production environment.
Q: Will idle-time tasks lose context across days? A: The official claim is that state, context, and decisions are persisted, surviving across sessions and days. Actual reliability must be validated on real long-running tasks; the specific persistence mechanism is subject to official documentation.
Q: Can ZCode 3.0 replace Claude Code? A: It depends on your scenario. Goals plus Subagents take a step forward on the autonomous delivery path, but three hard metrics -- model code generation quality, adaptation to your specific stack, and long-running task reliability -- must be validated on your projects. The Claude Code backdoor incident opened a window; whether ZCode keeps the users depends on the product.
References
- Sina Finance / Caiwen: Zhipu ZCode 3.0 upgrade coverage (2026-08-11)
- Yunbaba: ZCode 3.0 four-feature deep dive (2026-08-10)
- ZAKER, Zhihu, Tencent Cloud: ZCode 3.0 related coverage and discussion
- Related on this site: AI coding plan comparison | Crush coding agent open-source resource | AI coding context engineering SOP