Goal、Subagents、Remote Control与闲时任务四大功能上线

Core Highlights

ZCode, under Zhipu (GLM), today launched four new features at once: Goal, Subagents, Remote Control and idle tasks, and this capability is deeply optimized specifically for the GLM model family. Simply put, ZCode wants to upgrade itself from "an assistant that writes code" into "a programming operating system that can break down goals, dispatch sub-agents, stand guard remotely, and even make smart use of idle compute." In the Z.ai Code Bench test, GLM-5.2 paired with ZCode achieved an overall task pass rate 2.39% higher than when paired with Claude Code. That margin may sound small in casual conversation, but inside a rigorous engineering benchmark it represents a real and repeatable gap rather than noise, and it gives GLM users a concrete reason to prefer the native toolchain built around their own model instead of reaching for a foreign alternative by default every single time they start a project.

Specific Capabilities and What Happened

Goal lets a user state a high-level objective and then ZCode breaks it into a multi-step plan and executes that plan on its own. Subagents split a large task across several parallel sub-agents that do not collide with each other, which shortens wall-clock time on big jobs considerably. Remote Control supports running tasks on a remote machine so that closing the local window does not interrupt the work at all. Idle tasks exploit periods when quota is not tight to do work automatically, reserving precious peak compute for urgent jobs that cannot wait. Taken together, the four features turn "one person watching one terminal" into "one commander leading an automated squad" that works while the human sleeps. The effect is that a single developer can now orchestrate far more work than before, because the mechanical parts of planning, splitting, and waiting are handled by the system instead of by human attention that is easily exhausted by long and boring build loops.

Technical Details

Officials disclosed that ZCode's cache hit rate exceeds 98%, which means a large amount of repeated context does not need to be recomputed, making it both faster and cheaper at the same time. Combined with a 1.5x limited-time quota bonus, overall GLM Coding Plan usage approaches 1.8x the regular quota, a clear signal that users are indeed throwing heavier work at it than before. The optimization for GLM shows up in tool calling and long-context cooperation, where the model's strengths are used rather than fought against by the surrounding harness. High cache efficiency is not cosmetic: in agentic coding the same files and symbols get read again and again, so a good cache turns what would be a quadratic cost into something closer to linear, which is exactly the kind of engineering detail that decides whether a tool feels snappy or grinds to a halt on a real repository of meaningful size.

Comparison with Competitors

The direct rival is Claude Code. The 2.39% pass-rate advantage looks modest, yet in an engineering benchmark it already counts as a solid gap; more importantly, ZCode turns "tuned for GLM" into a selling point, giving domestic model users a more natural default choice instead of always bending to fit overseas tools. Where Claude Code is built around a specific foreign model and its idioms, ZCode is shaped to match how GLM actually behaves, which reduces friction for teams already invested in that stack and its conventions. The competition here is not only about raw scores but about which toolchain feels like home, and for a growing number of Chinese engineering teams that home is increasingly a domestic model with first-class local support rather than a product designed on another continent with different priorities and assumptions baked in.

Industry Impact or Applicable Scenarios

For small and mid-sized teams building on GLM, ZCode lowers the bar for "agentic programming" even further: goal decomposition, parallel sub-agents and remote execution all work out of the box without custom glue code. Idle tasks help budget-limited teams max out their compute utilization instead of letting quota sit idle overnight when nobody is watching the dashboard. The direction it points to is clear: programming assistants are moving from "completing code" toward "hosting the entire development flow" end to end. When planning, execution, monitoring and off-peak scheduling live in one place, the human role shifts from typist to reviewer, and that shift is where most of the productivity gain actually comes from in practice. For organizations worried about vendor lock-in, a GLM-native path also means they can keep their tooling inside an ecosystem they can negotiate with, audit, and trust with internal code safely. As coding agents move from novelty to daily infrastructure, the teams that win will be those who can trust the whole stack, and a locally tunable, GLM-optimized harness makes that trust easier to justify to security and finance alike. The release also signals a broader push by Chinese model labs to own not just the weights but the developer experience built on top of them.