# Coding agent

> A tool that reads a codebase, decides what to do, and writes code. Which ones matter, and why the differences change what gets installed.

Source: https://armature.tech/library/coding-agent
Published: 2026-09-03
Publisher: Armature, Inc. (https://armature.tech)

---

> A **coding agent** is a tool that reads a codebase, decides what to do about a request, and writes and runs code. Claude Code, Codex, Cursor, Gemini CLI, Cline, Aider and OpenCode are examples.

The defining feature is that it acts. An autocomplete suggests the next line. An agent takes a goal, explores the repository, picks an approach, edits many files, runs commands, and comes back with a result.

## Why they matter commercially

Because they buy things.

Given "I need a real database for this, something hosted", an agent will read the project, pick a vendor, install the client library, write a schema and add a connection string. A person often approves that diff without reviewing the choice of vendor.

So for a growing share of developer tool purchases, the buyer is an agent. It reads differently from a person, searches differently, and cares about different things. See [how coding agents choose tools](/library/how-coding-agents-choose-tools).

## They are not interchangeable

This is the part that surprises people. Across 5,292 judged sessions in 18 categories:

| Difference | Claude Code | Codex | Cursor |
| --- | --- | --- | --- |
| Ran a web search on decision tasks | 1.6% | 53% | not measured separately |
| Wrote its own implementation | 19% | 10% | 10% |

And they disagreed on the category leader in 10 of the 18 categories. Email: Postmark, Resend, Resend. Sandboxes: E2B, Modal, E2B. Authentication: WorkOS AuthKit, Auth0, Auth0.

A vendor measuring one agent is seeing part of the market.

## The parts of a coding agent

Three things, often confused.

**The model.** Claude Opus, GPT-5, Grok. The thing that reasons.

**The harness.** The program around the model: how it reads files, which tools it has, how it plans, what it remembers. See [harness](/library/harness).

**The environment.** The repository, the shell, the network, the permissions.

Two agents on the same model behave differently because the harness differs. Two harnesses on different models differ more. Both matter, and in our data which agent was running mattered more than most things a vendor controls.

## Where they run

| Shape | Examples |
| --- | --- |
| Command line | Claude Code, Codex, Gemini CLI, Aider, OpenCode |
| Inside an editor | Cursor, Cline, Copilot agent mode |
| Hosted, asynchronous | Cloud agents that open a pull request |

The shape changes what the agent sees first. An editor agent has the project open and indexed. A command line agent starts by listing files. A hosted agent works from a checkout and a task description.

## What this means if you sell software

Measure per agent, always. Match the lever to the agent: content reaches the agents that search, and the repository reaches all of them. Then check the numbers again after every update, because these tools change quickly and a result from six months ago is history rather than a fact.

## Common questions

### What is a coding agent?

A tool that reads a codebase, decides what to do about a request, and writes and runs code. Claude Code, Codex, Cursor, Gemini CLI, Cline, Aider and OpenCode are examples. The defining feature is that it acts rather than only suggests.

### How is a coding agent different from an autocomplete?

Autocomplete suggests the next few lines inside a file you are editing. A coding agent takes a goal, explores the repository, decides on an approach, edits many files, runs commands, and reports back.

### Do coding agents choose which products to install?

Yes, routinely. Given a one-line request such as needing a hosted database, an agent will pick a vendor, install the client library and write the integration, often without a person comparing options.

### Which coding agents behave most differently?

Claude Code and Codex. On decision tasks Claude Code ran a web search in 1.6% of runs and Codex in 53%, and Claude Code wrote its own implementation in 19% of runs against Codex's 10%.

## Read next

- [How coding agents choose tools](https://armature.tech/library/how-coding-agents-choose-tools) (Markdown: https://armature.tech/library/how-coding-agents-choose-tools.md)
- [Do Claude Code, Codex and Cursor pick the same tools?](https://armature.tech/library/do-claude-code-and-codex-agree) (Markdown: https://armature.tech/library/do-claude-code-and-codex-agree.md)
- [Harness](https://armature.tech/library/harness) (Markdown: https://armature.tech/library/harness.md)
- [Agent discoverability: the complete guide](https://armature.tech/library/agent-discoverability) (Markdown: https://armature.tech/library/agent-discoverability.md)
- [What about Gemini CLI, Windsurf, Cline and the others?](https://armature.tech/library/what-about-the-other-coding-agents) (Markdown: https://armature.tech/library/what-about-the-other-coding-agents.md)

---

Armature helps software products get discovered and used by coding agents.
Service: https://armature.tech/discoverability · Results: https://armature.tech/leaderboards/sectors · Contact: contact@armature.tech