Introduction

AI coding agents are tools that can read your codebase, plan changes, edit files, run commands, inspect outputs, and explain what they did. The important shift is not that they “write code from prompts.” It is that they can participate in a real development loop.

What They Are Good At

• Exploring unfamiliar codebases and summarizing key patterns quickly.
• Making repetitive edits across multiple files.
• Writing or updating tests once the desired behavior is clear.
• Drafting migrations, refactors, scaffolding, and internal documentation.
• Translating high-level intent into a first implementation pass you can review.

Where They Commonly Fail

• Inventing APIs, file paths, or framework conventions that do not exist.
• Choosing broad rewrites when a focused diff would be safer.
• Handling edge cases poorly when the task is underspecified.
• Reporting confidence that is not matched by real verification.
• Copying patterns from adjacent code without understanding the deeper constraint.

The Working Mental Model

Treat a coding agent like a fast collaborator with partial context, not an autopilot. The better you define scope, constraints, and verification, the better the output gets. When the task is ambiguous, ask for a plan first. When the task is risky, keep the change small and insist on evidence.

How To Use This Guide

Read the Getting Started section in order the first time:

1. Learn how stable context works.
2. Learn how to ask for code work clearly.
3. Learn how tools, MCP, and permissions change what the agent can do.
4. Learn how to verify work instead of trusting appearances.

After that, use the individual pages as reference links for specific topics.

A Good Default Workflow

1. Give the agent stable project context.
2. Ask for a plan if the change is non-trivial.
3. Ask for tests, docs updates, and explicit verification.
4. Review the diff, not just the explanation.
5. Use automation and CI to catch anything the agent misses.