Running coding agents headless, and where typed-decision models fit

We met on 29 September 2026 for the fourth and last office hours of the AI dev tools course. One member joined to talk about homework 4, running agents headless, and choosing models. The host showed how he runs ZCode on a remote machine through a Codex front end. We also discussed where a typed-decision model like Jev is likely to stay useful after the hype, and what comes after the course.

Homework 4: give the on-call agent a headless runner

Homework 4 asks you to build an on-call agent that answers an alert automatically. That agent needs a coding agent it can call in headless mode, without a person at the keyboard. The webinar used Claude Code's headless mode for this. The member did the homework in ZCode and asked it to find a headless-capable agent on their computer.

• ZCode found three options: Claude Code, OpenCode (installed as a VS Code extension), and ZCode's own built-in headless agent.
• Claude Code was logged out, so it couldn't answer automatically. Log in to any CLI agent before you call it from an unattended flow.
• The member ended up using OpenCode. The host suggested trying ZCode's own headless agent next.

Run ZCode through its headless mode

Z.ai gives you about 50% more tokens when you use its models through ZCode rather than directly. ZCode also runs token promotions. The host said he claimed a new 1-million-token promotion on the day of the call.

ZCode was recently open-sourced after an incident where it was found uploading data from users' computers to Alibaba Cloud. The company said it opened the code to win back trust. The discussion also mentioned that Grok did the same after a similar finding.

The host runs ZCode on a remote machine that has no desktop:

1. ZCode has a headless mode, so it can run without its desktop app.
2. zcodex is a Codex CLI fork that uses ZCode as its model backend. It keeps its configuration in ~/.zcodex, so it doesn't conflict with a normal Codex installation.
3. You give the task to the Codex interface on the remote machine, and ZCode does the work under the hood. The promotions still apply.

Before ZCode was open-sourced, the host had an agent reverse-engineer ZCode's minified JavaScript to understand the headless protocol. After the release, he pointed the agent at the source. He asked it to check how the desktop app talks to headless mode and to borrow what it could. That analysis led to several improvements in zcodex.

Organize parallel sessions with PocketShell

PocketShell Desktop is the terminal the host uses to manage many agent sessions on his remote dev box. About 95% of its code was written with ZCode.

• It groups running sessions by folder and shows each session as a tab, so you don't have to search for the right one in a terminal.
• You can connect to the same sessions from your phone.

The member runs an agent on an always-on Oracle Cloud A1 instance, which works like the host's remote computer. They're still working out how that agent should delegate coding tasks to coding agents. For homework and the capstone, they use ZCode because it's simple.

Models the host uses day to day

The host shared his current model choice.

• The Flash model on Z.ai is currently the host's favorite for many tasks. It's cheap and it can process images.
• It's a little weaker than the full GLM-5.3 model. It still handles most everyday work, like a Django or Python app or a small Rust program for Linux.

Typed-decision models after the hype

The member asked whether Jev, the typed-decision model from the previous week's workshop, will outlast the hype.

The host thinks two uses are likely to last:

1. Evals. Jev returns a class with a probability for each question instead of generating text, so it can replace an LLM judge. The host has ported some of his evaluations to Jev and uses it through OpenRouter. He found it cheaper and faster, because one request can check several criteria in a single call. He estimated it at about three times cheaper. He added the workshop as a bonus module in the new AI Engineering Buildcamp cohort.
2. Computer use. Current agents in ZCode or Codex control your computer through screenshots. The LLM analyzes each screen as an image, decides on an action, and moves the cursor, which is slow. Demos of Jev and of fine-tuned open models for this task look much faster.

The host found a hosting provider for Laya, the open alternative from the workshop, but hasn't used it since.

Try it yourself

Use these steps to act on the call without watching it.

1. Finish homework 4 with a headless agent. Log in to the CLI agent first, or try ZCode's built-in headless agent instead of a third-party one.
2. If you work on a remote machine, try zcodex to drive ZCode through a Codex interface.
3. Try PocketShell Desktop if you run several agent sessions at the same time.
4. Work through the Jev for LLM Evaluations workshop. Compare a Jev judge with your current LLM judge on the same eval run and record the cost difference.
5. For your capstone, start with one project and scope it to finish on time. The member plans to rebuild last year's arcade project with an automated team of agents instead of orchestrating the agents by hand.

After the course

This was the last office hours of the course. If you're stuck, post your question in Slack and the host will aim to reply within a day, possibly with a voice message. He's also planning a follow-up course for November that continues the AI dev tools course, with more focus on building a product.