Four Years of Coding with AI: From Chatbots to an Agent Fleet

Summary

This personal account traces the author’s shift from manually writing every line of code in 2022 to supervising a distributed fleet of coding agents in September 2026. ChatGPT first served as a search and explanation tool, followed by Phind, Perplexity, and Kagi, but the workflow remained a manual clipboard loop. GitHub Copilot moved assistance into VS Code through inline completions, while Ollama enabled local models for private work, prompt experiments, and tooling. The author built Harbor to start local backends and services, then tried direct-editing agents including Aider and Roo Code; these exposed a gap between agent interfaces and the models available at the time. Stronger models later made Claude.ai, Copilot chat, and Copilot agent mode useful for handling whole files, cross-file edits, and diff-based review. Remote access through RustDesk, Tailscale, SSH, and Zellij allowed work to continue from a phone and shifted inference to an always-on Strix Halo devbox. The author then organized skills, agent personas, notes, and infrastructure in lifeos, and created RUG and DRUG to delegate software work through layered SWE and QA agents. In 2026, Claude Code’s terminal-based workflow replaced those orchestrators, while the mi harness and shared skills repository made agent knowledge reusable across tools. OpenCode, Hermes, Cursor, Droid, Grok CLI, and Codex turned model choice into a scheduling problem governed by capabilities and quotas. Monitoring tools indexed thousands of sessions, and cloud runs enabled six-hour, timeboxed tasks with subagents. By September, Kandev assigned tasks to agents in separate worktrees across four Tailscale-connected machines, while a fleet CLI synchronized skills and checked host health. The author’s conclusion is that the main change was not simply model quality, but the location of the development loop: from the programmer’s hands to the editor, terminal, task board, and finally a managed fleet.