Let me share how I set up my Harness engineering environment using seven models. The big window is split into two parts, with Codex on the left and Zellij on the right, unified session window.
Codex is something everyone is familiar with, while Zellij is a terminal module with built-in multi-terminal management, allowing you to open many independent terminals for unified memory management. The next time you start, a specified session command can be invoked to restore all previously opened terminals and layouts, automatically recovering context (it's a better alternative to tmux, which can be a pain to get into. The advantage of Zellij is that you can use the mouse to hit different terminals and automatically copy selected text to the clipboard).
In terms of models, I'm using 7 models: Gpt5.5, Grok4.3, Gemini 3.1, Glm5.1, Composer2.5, Kimi (I subscribed to Claude’s $20 membership, which is pure scam and not durable at all, only used to bounce ideas with Gpt during critical times).
The division of labor is as follows:
1️⃣ Framework and product control:
Gpt5.5: The chief designer, responsible for the framework, task distribution, final review, and decision-making.
Grok4.3: Responsible for the first version review, raising questions from developers, and providing Gpt with the final review preview.
2️⃣ Developers: Gemini, Glm5.1, Composer2.5, Kimi2.6. They develop tasks in parallel according to the task flow document organized by Gpt.
After each development round, Grok conducts a unified first version review and gives feedback to Gpt for the final review. If there are bugs, Gpt will give specific model instructions for rework. If the acceptance passes, it moves on to the next round of task flow layout and distribution 🔄
All these processes, context recovery, integration, and development standards are documented in the Agent documentation, serving as the Harness engineering environment. It integrates the development standards from andrej-karpathy-skills/CLAUDE.md.
Claude Code has an ultracode automatic dynamic workflow mode. Claude autonomously assesses task complexity, automatically orchestrating parallel sub-agents to achieve multi-stage processes such as "understand → change → verify" without needing user-triggered execution.
In the future, we can try letting cc take over the entire process as the control brain, conducting different rounds of development. As long as the boundaries of Harness's specifications are clearly defined and personal demands and designs are articulated in advance using the Plan mode, it runs like a token slaughterhouse 🐷
#BTC #vibecoding