Claude Code and the Agentic Engineering Loop

From L2 to L3: the technical cliff

L2 tools build prototypes. L3 tools ship production code. The gap between them is real engineering: security, authentication, data integrity, error handling. Most people stall here.

What L3 looks like

You define "done." The agent figures out how. You review outcomes, not keystrokes. This is the Supervisor → Agent relationship.

The agentic engineering loop

1. Understand: Agent reads your codebase, docs, and existing patterns
2. Plan: Agent proposes an approach before writing code
3. Execute: Agent writes code, creates files, runs tests
4. Verify: Agent checks its own work against acceptance criteria
5. Iterate: Agent fixes issues found during verification
6. Deliver: Clean PR ready for human review

Tools

Claude Code

CLI tool that reads your entire codebase, understands the architecture, and makes changes across multiple files. Works in your terminal alongside git.

Cursor

AI-native code editor. Understands your project context and suggests changes across files.

The key mindset shift

Stop micromanaging. Write clear acceptance criteria. Let the agent work. Review the diff, not the process. If you're watching every keystroke, you're still at L2.