AI coding agent: what it is and how it works

Coding agent

A coding agent, or AI coding agent, is an AI agent that writes and changes code in your project on its own, for example in the terminal on your computer: it reads code, edits files and runs commands. It gets its task described in plain words. Such agents include Claude Code by Anthropic, Codex by OpenAI and Cursor by Anysphere.

How a coding agent works

You start the agent in your project’s folder and describe the task in plain words, such as a bug fix or a new feature. The agent reads the files it needs, changes them and runs commands, tests for example, to check the result. In the terminal you start Claude Code with the claude command, Codex with codex and Cursor CLI with agent. All of them then run on your computer.

You decide what the agent may do on your computer in its own settings. In Claude Code and Codex, that is where you set when the agent must ask for your permission before it changes a file or runs a command.

A coding agent in featureloom

featureloom does not run the agent and does not connect to your code. It holds what people and agents write into it themselves. The agent runs on your computer and connects to featureloom over a remote MCP server. In the app, such a connected AI assistant is called a console, and it acts with the rights of the person who connected it. You connect Claude Code or Codex to featureloom with one connect command from the app, pasted in your system terminal, and the Cursor editor with a button from the app.

The agent takes its work from the plan board. A plan is a document on a kanban board. The command from a plan, such as featureloom realizer INV-14, gives the agent the prompt with your team’s process, the whole plan document and the documents of the delivered plans it directly depends on. The agent can also take the next ready plan itself, one whose predecessors are delivered. The server does not let it take on a plan that waits on another.

From there, the agent:

takes the plan on, and the card moves to the column marked in realization and shows the person and the console working on it;
adds progress notes and, when the plan leaves something open, asks in the plan;
reports the plan delivered, and the card moves to the column of delivered plans.

The card of a plan in realization also shows a short note when the console gives one itself, such as the computer, repository and branch.

The agent chooses neither the column nor the order. A person reviews the work in the app: they press Accept or Do not accept. After a rejection, the agent reads the verdict, goes back to work and reports the rework. The How featureloom works page shows a plan’s whole road.

Questions

How is a coding agent different from an AI agent?

A coding agent is one kind of AI agent. Like any AI agent, it works toward a goal on its own and uses tools to get there, but its field is a project’s code: it writes and changes it. Working with such an agent, where it writes the code while you describe the task and review the result, is called agentic coding.

Does featureloom run the agent for me?

No. Claude Code, Codex or Cursor runs on your computer and is paid for with its provider. featureloom gives it plans and prompts over MCP and shows its work on the board.

Does featureloom connect to my code?

No. The agent on your computer reads and changes the code. featureloom holds what people and agents write into it themselves, such as the plan document, progress notes and the Realization note.

Which agents does featureloom work with?

For Claude Code and Codex the app gives you a ready connect command, and for Cursor a button. You connect another MCP client by signing in through the browser, if it runs on your computer and supports OAuth sign-in.