🔗 Finding my way with AI coding agents

The author initially disliked AI coding assistants, finding inline autocomplete intrusive and distracting from their work. After exploring GitHub Copilot Pro agents in Zed, they realized the core problem: AI agents operate at the wrong abstraction layer, viewing static code rather than the running application.

Tidewave, an Elixir-based tool by José Valim, runs directly inside applications at runtime and lets agents see what developers see. The author's workflow involves connecting Tidewave to Copilot, using Gemini 3.1 Pro to implement features while watching the agent test changes in the browser, then manually refactoring the generated code in Zed before committing.

This approach resolves the abstraction layer problem and keeps the developer engaged with their codebase. The author maintains manual control over code quality and structure, avoiding the pitfalls of fully autonomous agents while still leveraging AI's capabilities effectively.