Which of the seven coding agents to run
For parallel work, three things decide it: whether the agent reports its state, whether approvals can be waived, and whether it reads .claude/skills as it stands.
Table of contents
Side by side
| Reports state | Waives approvals | .claude/skills | |
|---|---|---|---|
| Claude Code | finished + waiting | --permission-mode | read as-is |
| Codex | finished + waiting | none | mirrored first |
| Antigravity | not wired | --dangerously-skip-permissions | a different file |
| Grok | not wired | --permission-mode auto | read as-is |
| Muse | not wired | --yolo | read as-is |
| GitHub Copilot | finished only | — | — |
| Cursor | finished only | — | — |
Reading the table
Leaving sessions unattended for long? Take one that reports state. The others have to be visited to learn whether they are done.
Want it to keep moving? Take one whose approvals can be waived. Codex is the one that cannot.
Have .claude/skills worth carrying? Grok and Muse read that directory themselves. Codex mirrors it; Antigravity wants a file of its own per directory.
You do not have to choose one
Each cell picks its own. Long unattended work on Claude Code, careful work on Codex, in the same grid.
Related
日本語版: which-agent-to-use