Zhipu ZCode just dropped a major upgrade today: four new features — Goal, Subagents, Remote Control, and Idle Tasks — and the user count hit 1 million. The real story isn't the feature list. It's the shift in positioning: ZCode is no longer "an assistant that helps you write code." It's now "an Agent that ships complex projects on its own."
On August 11, Zhipu announced the ZCode overhaul (source). Alongside the upgrade, all GLM Coding Plan weekly quotas were fully reset at 13:00 Beijing time.
Four Features: The Agent Finally Runs the Whole Pipeline
This upgrade adds a full "goal management + autonomous execution" layer to the Coding Agent. Before, ZCode worked like a conversational coding assistant — you ask, it answers, you iterate. Now you give it a verifiable goal and it plans, executes, traces, and verifies — end to end.
Goal: Target-Driven Autonomous Loop
This is the centerpiece. Set a verifiable target — "refactor the user module and pass all regression tests" — and ZCode handles the rest:
- Auto-decompose the task
- Modify code
- Run commands, execute tests
- Evaluate completion based on file changes, command output, and test results
- If the target isn't met, loop again until it is
The Goal dashboard shows per-round progress, duration, and results. You don't need to babysit it.
Subagents: Decompose and Run in Parallel
Large tasks get automatically split into subtasks, assigned to different modules running in parallel. Cross-file, cross-module coordinated changes that used to require manual context switching are now handled by Subagents autonomously.
Remote Control: Progress Anywhere
Once a task is running, switch to other windows and do your own thing. Remote Control lets you check progress from anywhere. Built for long-running tasks that take hours.
Idle Tasks: Smart Scheduling
Non-urgent tasks don't compete with real-time calls during peak hours. They automatically execute during low-load periods — saves quota and avoids contention.
Performance: GLM-5.2 + ZCode vs. + Claude Code
Zhipu's internal Z.ai Code Bench shows GLM-5.2 + ZCode outperforming GLM-5.2 + Claude Code by 2.39% in overall task pass rate. (source)
| Metric | GLM-5.2 + ZCode | GLM-5.2 + Claude Code |
|---|---|---|
| Overall task pass rate | +2.39% | Baseline |
| Complex cross-file tasks | Gap widens | — |
| Cache optimization | GLM Coding Plan quota +~30% | — |
| Campaign boost | 1.5× through Aug 31 | — |
2.39% may sound modest, but this is an overall pass rate including both simple and complex tasks. The advantage concentrates on cross-file, multi-stage tasks requiring final verification — there, the gap is substantially wider.
Two reasons ZCode pulls ahead: cache hit rate optimization boosted Coding Plan effective quota by ~30% (more rounds for the same budget), and the Goal + Subagents decomposition algorithm was specifically tuned for the GLM family.
1 Million Users + Full Quota Reset
ZCode officially crossed 1 million users (source). As a milestone celebration, all GLM Coding Plan weekly quotas were fully reset on August 11.
Combined with the 1.5× quota campaign running through August 31, actual usable capacity nearly doubles:
| Tier | Normal Weekly Quota | Current (1.5× + cache boost) |
|---|---|---|
| Lite | ~400 prompts/week | ~780 prompts/week |
| Pro | ~2000 prompts/week | ~3900 prompts/week |
| Max | ~8000 prompts/week | ~15600 prompts/week |
GLM-5.2 / GLM-5-Turbo peak multiplier stays at 3×, but off-peak usage during the 1.5× campaign only burns at 1× instead of 2×.
Zhipu's Coding Ecosystem Comes Together
Zoom out, and this upgrade completes the third pillar of Zhipu's "base model + Coding Harness + subscription service" architecture. GLM-5.2 provides the model capability. The Coding Plan bridges compute to developers. ZCode closes the loop on Agent execution in real engineering environments.
Claude Code is more general and supports more models, but it's Anthropic's toolchain — it will never be deeply optimized for GLM. ZCode's tight GLM-family coupling is its moat.
Caveats
- ZCode currently only supports GLM-series models — no mixing with DeepSeek, Kimi, or other third-party APIs. For multi-model setups, Claude Code or Cursor remain more flexible.
- The 1.5× quota campaign and full reset are time-limited — they end August 31.
- Goal is fundamentally an autonomous Agent loop. Complex task success depends on model capability and task decomposition quality — real-world experience will vary by task type.
- If the Coding Plan cache storage policy changes in the future, the sustainability of the quota boost would be directly affected.