Cursor

An editor that is also the agent

Cursor is an editor built around a coding agent rather than an editor with one bolted on. Its own description is "a coding agent for building ambitious software", and the documented work is understanding a codebase, planning features, fixing bugs and reviewing changes.

That puts it beside Windsurf rather than beside Codex or Claude Code, and the split is about where you are looking. A terminal agent asks you to read a diff it produced. An agentic editor keeps you in the file while it works.

Three interaction modes, doing different jobs

The documentation names Tab, Agent and Composer, and treating them as one feature is the usual mistake.

Tab is predictive editing. You are typing, and it proposes the rest, including edits elsewhere. This is the feature with no equivalent in a terminal agent, and the one that changes how the hour feels rather than what you can accomplish.

Agent takes a task and works across files.

Composer is the multi-file construction surface, and Composer 2.5 also appears as a selectable model, which is worth knowing because the same word names both a mode and a model.

Plan Mode sits in front of the building work, producing an approach before changes land. The same instinct as reviewing a plan before a terminal agent starts.

Model choice is the real differentiator

You pick the model. Claude Opus, Sonnet and Haiku variants, OpenAI's GPT-5 series, Google Gemini versions, and Cursor's own Composer models.

Two consequences. You can put a stronger model on a hard refactor and a cheaper one on routine work, which is the discipline choosing the right model argues for, applied to your own tooling. And a proprietary model in the list means some of what the product does well is not portable, so a comparison run on Composer tells you about Cursor rather than about the underlying models.

Customisation, in one place

Plugins, skills, MCPs and rules are configured together. Rules are the project-level instructions equivalent, the same idea as a CLAUDE.md or an AGENTS.md, and MCP support means it reaches your own systems through the protocol MCP in one sitting describes.

Set the rules up on day one. An agentic editor without project instructions spends its first move rediscovering conventions your team settled years ago, and it will guess differently each session.

What the documentation did not tell me

Two things a team evaluating it will want, which the overview does not cover: how the codebase is indexed and retrieved, and whether there is a CLI or headless mode for CI.

Both matter. Indexing decides what happens on a large repository, and a headless mode decides whether the tool can review pull requests as well as help you write them, which is where Codex and Claude Code have a documented answer. Check the current docs rather than assuming either way.

Where it fits

Reach for Cursor when the work is hands-on: you are in the code, you want completion that anticipates, and you want to watch changes land in the file.

Reach for a terminal agent when the work is a task you want done while you do something else, or when it has to run in CI. Several teams run both, and the two are not really competing for the same hour.

Reading documentation for something moving this fast

Windsurf makes this point and it applies here equally. Mode names, model lists and pricing in this category change within months. Composer's dual role as a mode and a model is exactly the kind of detail that gets renamed.

Treat the interaction shapes as the durable part. Predictive editing, an agent across files, a plan step before changes, and project rules are the concepts; what they are called this quarter is not.

Further reading

• Cursor, Documentation: the modes, the model list, and the customisation surface.
• Windsurf: the neighbouring agentic editor.
• Claude Code and Codex: the terminal alternatives.
• Choosing the right model: the same decision, applied to your tooling.