Build1 publisher2 min readPublished
Mistral's Vibe keeps its approval classifier on a Mistral model even when GLM 5.3 writes the code
Mistral has added Z.ai's GLM 5.3 to Vibe, its terminal coding agent, and moved the smart-approve classifier onto a fast Mistral model whatever model is active. Teams that pick GLM 5.3 to write code still get a Mistral model deciding when a person must approve an action.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Vibe's configuration documentation lists GLM 5.3 as a preset in its provider and model list.
- The release notes also list improvements to smart approve's pre-checks, escalation and mid-turn reclassification.
- According to the changelog, smart-approve behavior persists when a user switches modes mid-session, and approvals across subagents have changed.
- Mistral's AGENTS.md documentation describes how subagents are declared and used alongside regular agents.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Teams can change which model writes their code, but on the changelog as reported they cannot change the model that decides when a person gets asked to approve.
- decision Whether to use GLM 5.3 over other Vibe models has to be settled on each team's own code, because the documentation does not establish how it compares.
- cost The workflow design falls on the adopting team: which tasks get agent help, where approval is required, and which model gets tested on which work.
Smart approve is the step in Vibe that handles the moments when an agent action needs a user's approval [12]. According to the Vibe changelog, as summarized in a dev.to write-up, its classifier now runs on a fast Mistral model whatever model the session is using [5]. Switching a session to GLM 5.3 changes the code being proposed. The judgement about which actions reach a person stays with the same classifier [5].
I think the design is right for a team trialling a new model. Approval behavior can be tested once and held still while the worker model is evaluated. If the classifier followed the session model, a team swapping models would be re-testing its code quality and its gate at the same time. The trade-off is ownership. Teams pick the model that writes the code and decide how work is routed. The model that classifies each action is Mistral's in every session, as the changelog is reported [5].
A fixed gate matters most once work is split up. Mistral's announcement describes Work and Code modes and says custom modes and subagents can route specialized work inside one session [8]. In a multi-step task the agent may move between planning, implementation and specialized work without the user rebuilding approval preferences each time, according to the dev.to write-up [13].
GLM 5.3 is a hosted, third-party open-weights model from Z.ai that Mistral lists on its own platform [2]. Mistral describes it as designed for long-context coding and agentic workflows [3]. That is a design intent. A benchmark from someone else's workload would carry over only if its tasks looked like a team's own repositories, prompts and review requirements. The dev.to author advises testing GLM 5.3 against exactly those [15].
None of this removes review. The author writes that code changes, commands and outputs still need an approval approach suited to the project and its risk level [11].
What to watch
- A Vibe setting that lets teams choose the smart-approve classifier model, or a statement from Mistral that it stays fixed.
- Published comparisons of GLM 5.3 against other Vibe-compatible models on coding tasks.
- Detail on how approval preferences pass from a parent agent to subagents declared in AGENTS.md.