The flag that stops each coding agent asking for permission
You give a coding agent a twenty-minute job, go and make coffee, and come back to find it stopped at minute two, waiting for you to approve a file read. One agent doing that is an annoyance. Several doing it is a fleet that does not run.
Nearly every command-line agent has a way to stop asking. The names are different everywhere, and they are not all the same size of promise.
The flags
These are the options DevThrottle adds for its recommended setting:
- Claude Code:
--permission-mode auto - Codex:
--dangerously-bypass-approvals-and-sandbox - Gemini:
--yolo - Cursor:
--force - Grok:
--always-approve - GitHub Copilot:
--allow-all
Two of the agents DevThrottle supports have no such option to give. Pi's approval flag only trusts files inside the project, and OpenCode keeps permissions in its own configuration file rather than on the command line.
Claude Code also has --dangerously-skip-permissions. It is available as its own named choice, and it is not the one DevThrottle recommends.
Read the names
The names tell you how much each one gives away. Codex's flag says it in full: approvals and the sandbox both go. Treat every flag on this list as what it is, an instruction to the agent to act without checking with you first, on whatever that user account can reach.
When it makes sense
This is judgment rather than product fact, so here is ours. Running without prompts is reasonable when a mistake is cheap to undo:
- The agent has its own copy of the code. A worktree on its own branch means the worst day is deleting a folder. See two agents in one folder do not fight loudly.
- Nothing lands unreviewed. The check moves from every tool call to the finished change, which is a far better use of your attention. See the summary is a claim, the diff is the evidence.
- Credentials are out of its reach. See your agent needs the password, it does not need to see it.
It makes less sense on a machine holding things you cannot get back, or in a folder with no version control behind it.
How DevThrottle handles it
A freshly detected agent starts on the recommended setting above, because a fleet that stops on every tool call does not run. Nothing is forced. Standard, the agent's plain command line asking for permission as it normally would, can be selected for all eight. You change it per agent under Settings, Agents, where the editor shows the exact command that entry will launch with as you edit it.
DevThrottle does not modify the agents, so the same flags behave as they do in your own terminal. The full table is on Supported agents, and the settings live on Agents.