How Does an AI Coding Agent Know What Code to Change?

You tell an AI coding agent: "Add dark mode to this application." It doesn't magically understand your entire codebase — it has to figure it out.

A simplified workflow looks like: Developer Request → Repository Discovery (which files matter?) → Context Retrieval (what existing patterns should it understand?) → Planning (what needs to change?) → Tool Calls (read files, search code, edit files, run commands) → Tests (does the change actually work?) → Feedback (if something fails, inspect, modify, test again) → Final Change.

That's fundamentally different from Prompt → LLM → Code. The model isn't operating in isolation — it's sitting inside an orchestration loop, and that creates some interesting architecture problems. What happens when the agent retrieves the wrong files? What if it misunderstands the existing architecture? What if a tool call fails? What if the generated code compiles but breaks a critical workflow? What if the agent keeps trying the same fix?

That's why a production coding agent needs more than a powerful model. It needs context management, tool permissions, execution limits, validation, tests, observability, recovery, and human review.

There's an important distinction: The agent can propose a change. The system should verify the change.

That's why tests and validation become even more important as AI writes more code. The interesting future isn't "AI replaces software engineers" — it's "software engineers become the architects of systems where AI performs parts of the implementation loop." The hard engineering problem shifts from "can the model write code?" to "can we build a reliable system around the model that knows what context to use, what actions are allowed, and how to verify the result?"

That's a much more interesting architecture problem.

Would you let an AI coding agent merge directly into production without human approval?