Coding agents

Agents that write, review, fix and ship code from a task description, inside a repository or an editor.

18 agents, each with its inputs, outputs and review level checked by a person.

Coding agents take a task in plain language and do the work inside a real codebase: they read the repository, plan the change, write code and tests, run them, and open a pull request or apply the change in your editor. They differ from autocomplete in that they act across files and over several steps, and they differ from one another mostly in where they run (a cloud sandbox, your terminal, your editor) and in how much they do before a person looks. The listings here record what each agent needs connected, what it hands back, and whether a person reviews the result before it is merged.

Choose by the shape of your work. For a backlog of small, well-specified tickets, a cloud agent that opens pull requests lets several run at once. For exploratory work, an agent in the terminal or editor keeps you in the loop at each step. For a new project from nothing, the app-building agents produce a running site from a description. In every case, look at the setup level: some need only a repository connection, others need a developer to install a runtime.

The agents

Questions

What is the difference between a coding agent and code completion?

Completion suggests the next few lines while you type. A coding agent takes a whole task, works across files, runs commands and tests, and returns a change or a pull request. The listings here are agents in that sense.

Do coding agents merge code on their own?

Most open a pull request or propose a change for a person to review, which the listing marks as human review required. A few can be configured to merge or deploy without review; their listing says so under the review level.

How is pricing usually structured?

Per seat per month, per task or by compute used, and often a free tier with limits. Each listing shows the model the vendor publishes and the date the price was checked.