Don't Use Subagents on ZCode (GLM 5.3 Review)

We tried GLM 5.3 on ZCode and built 16 projects with it. The model itself is strong, but the subagent setup in ZCode let us down. Here is what we found and how to use GLM 5.3 well.

The model is much better

GLM 5.3 uses the same base model as GLM 5.2, but it got a 50% coding improvement from post-training. That is a real jump. On Terminal Bench 3.0 it sits slightly behind Fable 5 and GPT-5.6, ahead of Kimi K3. On automation benchmarks it even leads Kimi K3. Z.AI also says it is state of the art on CyberJim for vulnerability discovery, which matters for security work.

So the raw coding ability is there. The problem is how we used it. We ran every project through subagents to save tokens and time, but ZCode's documentation says project memory applies to the main agent, not subagents. We burned 40% of our weekly limit on the light coding plan in one day. The subagents built all the code, but the results were disappointing.

Subagents on ZCode are weak

We ran the same 16-project visual benchmark suite we use for other models. With subagents, the outputs were messy. The city scroll website was incoherent because we gave it no reference. GLM 5.2 did great when we fed it the Astro site specs and UI, but this time it was 100% freestyle. It still scrolled smoothly, but it lacked direction.

Some games worked but felt off. Ember Glider had a weird shaking penalty when you miss a ring, and the movement was hard. Jabberwock had a strange camera where pressing S moved you forward. The tower defense game would not let you place turrets on pads. Neon Drift and the Universe Simulator were just black screens. The ancient Chinese architecture build looked wrong and floated in the air.

A few things stood out as decent. The mechanical watch simulator had nice lighting details. Vice City looked different from other models, more like an old Nokia game than the usual neon vibe. The office life sim had usable pathfinding for chairs. But these were exceptions.

The petri dish test was interesting. Predators sometimes died while prey survived, so GLM 5.3 seems to favor the underdog. That is not a bug, just a bias in how it thinks.

Use the main agent, not subagents

If you use GLM 5.3 on ZCode, tell it to code the project itself. Do not have it coordinate with subagents. ZCode's harness is not built for good subagent work. Claude Code likely does subagents much better because the harness is designed for it. We plan to test GLM 5.3 again with the main agent doing everything, one project at a time. That will take longer but should show what the model can really do.

GLM 5.3 has no image capabilities, so visual tasks need guidance and references. It is text-only, and our benchmark runs rendered graphics locally on GPU. That means you need to be specific about what you want.

The prompts for all 16 projects are on our website, superbash.ai, if you want to try it yourself. The TL;DR is simple: GLM 5.3 is good at coding, but ZCode subagents are not. Build with the main model and set up your own loops.