4  What Is Agentic Coding?

4.1 The idea

You’ve probably used AI chat — you type a question, you get an answer. Maybe you’ve used autocomplete in an editor, where an AI suggests the next few characters as you type.

An agentic coding tool is different. It’s a program — sometimes called an agentic harness — that gives an AI model access to your actual development environment: your files, your terminal, and your git repository. Instead of just answering questions, it can:

  • Read your files to understand your project
  • Search across your codebase for patterns or definitions
  • Edit files directly
  • Run commands in your terminal (with your permission)
  • Iterate — check its own work, fix mistakes, and keep going

4.2 The agent loop

When you give an agentic tool a task, it doesn’t just generate text. It runs a loop:

  1. Think about what to do next
  2. Act — read a file, run a command, make an edit
  3. Observe the result
  4. Repeat until the task is done (or it gets stuck and asks you)

This is why they’re called “agentic” — the model is acting on your behalf, making decisions about what tools to use and in what order.

4.3 Examples of agentic harnesses

There are several tools that work this way. In this clinic you can use any of them through the shared gateway:

  • Claude Code — Anthropic’s CLI tool, built for Claude
  • OpenCode — open-source, works with Claude and open models
  • GitHub Copilot CLI — GitHub’s coding tool
  • Gemini CLI — Google’s coding tool

They all follow the same basic pattern: you give a prompt, the agent works through it using tools, and you review the result. The commands and interface differ, but the concept is the same.

4.4 What an agent is not

  • It’s not magic — agents make mistakes, especially on complex or ambiguous tasks.
  • It’s not a chatbot — keep interactions focused on tasks, not open-ended conversation.
  • It doesn’t remember between sessions — that’s why you use handoff notes.
  • It needs verification — see Checking the Agent’s Work.