The AI Coding Agent Stack

Three CLI agents, three different strengths: Claude Code (Anthropic) is the most capable for agentic multi-file tasks and has the best judgment on when to ask vs act — best for complex features and refactors. Amp (Sourcegraph) is built for large codebases with deep repo indexing — best for monorepos and navigation-heavy tasks. Aider is open-source with strong Git integration — best for budget-conscious setups or teams that want full control. Pick one as your primary and install it first.

For Claude Code: install via `npm install -g @anthropic-ai/claude-code`, set ANTHROPIC_API_KEY, run `claude` in your project root. Claude Code reads your CLAUDE.md for project context — create one with your stack, conventions, and test commands. For Amp: install from amp.dev, authenticate with your Sourcegraph account, and let it index your repo. Both support multi-file edits, terminal commands, and test execution within the agent loop.

Install Aider: `pip install aider-chat`. Run `aider --model claude-3-5-sonnet-20241022` to use Claude as the backend, or `aider --model gpt-4o` for OpenAI. Aider excels at focused, file-specific edits with explicit Git commits per change — use it for smaller, well-scoped tasks where you want atomic history. Pair Aider with Claude Code: use Aider for surgical edits, Claude Code for larger agentic tasks.

Install Cursor from cursor.com — it is a VS Code fork with AI deeply embedded. Use Cursor for inline completions, quick edits with Cmd+K, and the Composer for multi-file changes within the IDE. Cursor complements your CLI agent: the CLI agent does the heavy agentic lifting outside the editor, Cursor handles fast inline work while you are actively coding. Configure Cursor to use Claude 3.5 Sonnet or GPT-4o in Settings → Models.

AI coding agents ship code fast but can introduce subtle regressions. Establish three guardrails: (1) always run your test suite after each agent session before committing; (2) use Git branches for every agent task — never let agents commit directly to main; (3) create a short AGENTS.md that lists off-limits files, required test commands, and code style rules — both Claude Code and Aider respect this file. Review agent diffs the same way you review a junior engineer PR: check the logic, not just the output.