38. Building a Code-Change Agent Safely

High Stakes

An agent that can write code is an agent that can delete your database or introduce a backdoor. You cannot treat a coding agent like a chatbot.

The Safe Architecture

We introduce the Proposal Pattern:

1. Analyst Agent: Reads the issue and the code. Writes a textual PLAN.
2. Human: Approves the PLAN.
3. Coder Agent: Generates the diff based on the PLAN.
4. Reviewer Agent: Critiques the diff (security, style).
5. Test Runner: Runs the tests on the diff.
6. Human: Merges the PR.

Notice the two Human checkpoints. This is non-negotiable for production code.

Where to go next

• 38.1 The “proposal-only” agent pattern
• 38.2 The “diff generation” agent pattern
• 38.3 Test-driven agent behavior
• 38.4 Static analysis and lint gates
• 38.5 Security gates and dependency scanning