An Ex-Meta Engineer's Guide to Agentic Coding Setup
Ex-Meta/Microsoft/Atlassian principal engineer Kun Chen shares his terminal-centric workflow for managing AI coding agents instead of writing code himself.
Kun Chen, a former L8 principal engineer at Meta, Microsoft, and Atlassian who led Atlassian's Rovo Dev product, now works solo and has shifted from writing code himself to directing a fleet of AI agents like an engineering manager. He describes reaching a flow state where he ships more than 30 quality PRs a day, with his own judgment on what to build and evaluate now being the only bottleneck.
His setup is deliberately terminal-centric and vendor-agnostic: WezTerm as terminal emulator, Claude Code and OpenCode as agent harnesses, Neovim (with oil.nvim, neogit, and snacks.nvim plugins) for quick file and git review, and tmux for persistent, keyboard-driven session management across devices. He deliberately avoids agent-specific 'fancy' features like auto-managed memory, arguing they lock users into a single vendor when model performance leadership is still unsettled.
To make the workflow concrete, Chen walks through a real feature—image input—in his side project 'Hi Bit,' an AI tutor he's building to teach his son agentic engineering, tracing it from initial idea to merged pull request. He also touches on the growing role of voice input in prompting agents effectively, offering engineers a practical reference for structuring their own agent-driven development practices.