What is an AI coding agent?
A program that uses a language model to work toward a coding goal by itself: reading files, editing them, running commands, and deciding what to do next from the results.
Updated October 1, 2026 · Other tools described as of October 1, 2026
Short answer. An AI coding agent takes a task, not a line, and works in a loop: read, change, run, look at the output, repeat. That loop makes it useful and makes it risky, so what matters most is where it runs, what it may do without asking, and how you check its work.
Agent vs autocomplete
Autocomplete suggests the next few lines while you type. An agent takes a whole task, like "fix the failing test", and works in a loop: look at the code, change something, run a check, read the output, repeat. You see the result after several steps, not every keystroke.
What an agent needs
- Tools: listing, reading and searching files, editing them, and running commands.
- Limits: which of those tools it may use without asking. See autonomy levels.
- A place to run: your own machine, an isolated machine, or someone's cloud.
- A way to check the work: tests or other commands that pass or fail.
- A person at the end who decides whether to keep the change.
Where agents run
- Your machine
- ExamplesClaude Code or Codex in a terminal or IDE
- What it can reachWhat your user account can reach, minus what its permission settings block
- A machine of its own
- ExamplesA Docker container or VM, such as a Macaron Computer
- What it can reachOnly what you put on that machine
- The vendor's cloud
- ExamplesCodex cloud tasks, Cursor cloud agents, Copilot's cloud agent
- What it can reachThe cloned repository and any secrets you configure there
| Where | Examples | What it can reach |
|---|---|---|
| Your machine | Claude Code or Codex in a terminal or IDE | What your user account can reach, minus what its permission settings block |
| A machine of its own | A Docker container or VM, such as a Macaron Computer | Only what you put on that machine |
| The vendor's cloud | Codex cloud tasks, Cursor cloud agents, Copilot's cloud agent | The cloned repository and any secrets you configure there |
The same agent can run in all three. Claude Code, for example, runs locally, on Anthropic's cloud for web sessions, or inside a container you provide.
How agents go wrong
- Claiming work it didn't do, like saying tests pass without running them.
- Getting to green by weakening a test instead of fixing the code.
- Drive-by changes: renames and refactors nobody asked for, buried in a large diff.
- Running a command with side effects it didn't understand, like a migration against a real database.
- Reading a secret it was never meant to see, because it was in the environment.
Keeping an agent in check
- 01Run it where a mistake is cheap: an isolated machine, not the laptop with your SSH keys.
- 02Give it only the credentials the task needs, and scoped keys with spending limits.
- 03Have the system run your checks before and after, rather than trusting the agent's report.
- 04Keep each turn undoable, so a bad step costs one turn, not the session.
- 05Read the diff before you keep it.
In Macaron
Macaron runs Claude Code, Codex or its own agent, in a folder on your Mac or on a Computer, an isolated Linux machine. Each turn ends with a card showing the files changed, with an undo; your project's checks run before and after; an optional second model reviews the diff; and the receipt is built from what actually ran.
Questions
Is an AI coding agent safe to run?+
As safe as its limits. Run it where its commands can't reach anything valuable, give it scoped keys, and review its changes before keeping them.
Is Copilot autocomplete an agent?+
Inline suggestions aren't. Copilot's cloud agent, which takes an issue and pushes a branch, is.
Do agents need internet access?+
To install dependencies, usually. To do the task itself, often not. Codex's cloud tasks, for example, install with network access and then run offline by default.
Can an agent work on several tasks at once?+
Several agents can, each in its own workspace or machine so their changes don't collide.
What's the difference between an agent and an agent harness?+
The agent is the loop that decides what to do. A harness, like Macaron, is the environment around it: where it runs, what it can reach, and how its work is checked and kept.
Sources and dates
- Claude Code docscode.claude.com, read October 1, 2026
- Codex approvals and securitylearn.chatgpt.com, read October 1, 2026
- GitHub docs: Copilot cloud agentdocs.github.com, read October 1, 2026
Facts about other products come only from these pages, on the dates shown; prices are as published that day. Macaron's own details are as shipped on October 1, 2026. Products change, so check their sites for the latest.