Skip to content

AI agent tools

A very subjective comparison of the user interfaces of AI agent systems by parameters I find influencing whatever I am doing with them, which is mostly writing text or code.

Legend

Supported

Maybe not

© Agent Diff review comments Plan mode
Anthropic Claude Code (CLI)
Claude Desktop (Code)
SpaceX › Anysphere Cursor Editor
Cursor CLI
Google Gemini CLI
Antigravity Editor
Antigravity CLI
Zed Industries Zed
OpenCode Anomaly OpenCode OpenCode CLI
OpenCode OpenCode Desktop
OpenAI Codex CLI
Codex Desktop
Codex Remote

Diff review comments

A comment editor attached to a line in a diff

Whenever I review code an agent has written, and even more so when it is text, I see issues that are opportunities to learn. Every single time.

These issues can lead to new checks, linters, guidance, or skills. I want to comment on a specific line in the diff and give the agent feedback there. Often, I immediately ask it to /retro the issue: my skill turns a mistake into improved guidance.

Plan mode

A plan to conquer the world as a doughnut

Robert Martin argues that large specifications written up front bring back the problems of waterfall. I agree, and I no longer write specs. I do write plans quite a bit.

One reason is to split the work: a smarter agent can write a plan, and a team of lighter models can implement it. This resembles the orchestrator–workers pattern, where an agent divides work among other agents.