Prominent Developer Migrates Core Workflow to Linux, Citing AI Agent Performance and macOS Limitations

Previously a dedicated macOS user, the developer’s migration was primarily instigated by critical performance and efficiency issues when running AI agents, specifically Codeex, on even high-end Apple silicon machines like the M5 Max MacBook Pro. Intensive CPU utilization, often consuming over 50% of cores for single browser-based AI tasks and generating upwards of 30 processes per sub-agent due to macOS’s CIS policyd security mechanisms, rendered the development environment “unusable.” Further compounding these challenges, Apple’s APFS file system demonstrated significantly slower performance for common development tasks involving numerous small files, such as git clean and PNPM install, taking 35 seconds on a Mac compared to under 10 seconds on Linux. The inability to close a laptop while long-running agent tasks executed also imposed significant mobility constraints.

To address these limitations, the new workflow leverages a distributed architecture of on-premises Linux machines for compute-intensive tasks, accessed remotely via SSH, T-Mux for session persistence, and Tailscale for secure networking. A specialized tool, T3 Code, enables remote GUI-based development, restoring capabilities like screenshot pasting to agents. Out-of-band management of physical machines, including bootloader control and OS flashing, is achieved using network KVMs like the Glinaut Comet Pro, supplemented by robotic “Fingerbots” for remote power cycling. The developer’s MacBook now functions as the orchestrator, with agents capable of configuring and managing the Linux fleet. This setup has enabled longer, more complex agent-driven workflows, significant performance gains, and better resource utilization, while also circumventing the high costs of unsubsidized cloud AI agent hosting for personal projects. The developer noted that cloud offerings for AI agents, such as Codeex’s and Claude Code’s cloud products, remain challenging to set up and lack the control and functionality of a self-hosted, distributed environment.