AI coding agents: a shortlist for your first trial

A useful shortlist gives you a reason to try each candidate. Begin with a repository you know, a small change, and a way to check the result. Choose two candidates from the workflows below. Running the same task in both will tell you more about your fit than collecting another dozen product names.

This is an editorial starting set, not a benchmark ranking. Product descriptions come from official documentation checked on October 7, 2026. The suggested trials are ours; we have not measured comparative performance for this article.

Cursor: start beside the code

Cursor’s Agent can search a project, edit files, and run terminal commands from the editor. Consider it when you want to inspect the code while discussing a change. See the Cursor Agent documentation.

Try this: choose a small interface bug with a visible before-and-after result. Ask for the cause, the smallest relevant patch, and the checks that support it. Inspect both the diff and the rendered interface. Count the time you spend steering and reviewing, including manual repair.

Read the Cursor entry.

Kiro: try a specification before the patch

Kiro’s specs organize work into requirements or bug analysis, design, and implementation tasks. These artifacts give you something to inspect before code changes accumulate. That is a reason to try Kiro if your usual difficulty is deciding what a feature must do. See Kiro’s specification workflow.

Try this: describe a small feature with one awkward edge case. Review the acceptance criteria before implementation. Afterward, trace each criterion to a check or visible result. Record whether the specification reduced correction time enough to justify preparing and reviewing it.

Read the Kiro entry.

Claude Code: work through the repository and shell

Claude Code can read a codebase, edit files, and run commands. Its terminal interface is one way to use it; official documentation also describes IDE, desktop, and web surfaces. Consider the terminal workflow if your work moves between searches, test commands, and diffs. See the Claude Code overview.

Try this: reproduce a failing test in a clean worktree. Ask the agent to trace the failure before editing, fix the cause, and run the relevant checks. Review whether the patch solves the intended behavior or merely changes a test expectation.

Read the Claude Code entry.

Gemini CLI: another terminal candidate

Gemini CLI provides an agent in the terminal with tools for files and shell commands. It is a useful second candidate for a comparison that keeps your repository and shell workflow fixed. Read the Gemini CLI documentation for installation, authentication, and tool details.

Try this: reuse the same bug and starting commit as your other terminal candidate. Keep the brief, access, and time limit the same. Save the patch, command results, and correction history from both attempts.

Read the Gemini CLI entry.

Make the next decision smaller

Check that your account, operating system, repository, and data rules support the workflow. Read current plan limits before paying. Then run one ordinary case and one awkward case using the trial scorecard. Keep setup, waiting, and human attention as separate measurements.

If you want to hand off work while you are away from the editor, read editor, terminal, or cloud before expanding the shortlist. Browse the full agent list when you know which constraints matter.