Control the agent

Coding and planning

Coding mode exposes the normal enabled tool set under the active permission policy.

Planning mode changes instructions, tool availability, and enforcement. Research remains available; edit and write are constrained to Nerve’s plan directory; launch and background-process mutations and mutating Jira/Confluence operations are unavailable; shell commands pass a planning guard. A plan must be presented for review before implementation.

Select coding or planning mode in the agent controls; agents cannot change their own mode. In planning mode, the agent uses plan_mode_present to submit its plan. Accept & Implement approves the plan and transitions to coding mode; rejecting the plan keeps planning mode unchanged and leaves the agent idle until you send a follow-up prompt. You can also change mode directly with the user controls.

Planning restrictions reduce accidental mutation but are not an operating-system sandbox.

Permission levels

Read-only: permits local inspection and interaction, plus enabled sub-agent prompting and stopping. It denies other capabilities unless an applicable override changes that decision. It is not a restriction on a child’s profile-selected authority.
Supervised: automatically allows safe local reads and audited read-only integrations, then asks for commands, writes, opaque execution, and network access unless a matching permission rule allows the request.
Autonomous: executes allowed risks without normal approval pauses, while explicit block exceptions and hard guardrails still apply.

Policy is evaluated for every dispatch and recorded with the tool lifecycle.

Model and thinking

Select an authenticated model and one of its declared thinking levels. Nerve knows seven possible levels from off through max, but the picker only presents supported choices and runtime resolution can clamp a stale request.

Changing model, thinking, mode, or permission on an active conversation does not recreate it. Changes made during a run apply to subsequent work—especially the next provider request—not the request already streaming.

Delegation and authority

The built-in Read only rule set explicitly permits subagent_prompt and subagent_stop when delegation is enabled. A sub-agent’s profile selects its own rule set, either read-only or autonomous; the lead’s selection is not inherited as a ceiling. Autonomous sub-agents (such as Developer) work in the shared worktree without asking you, so enabling them grants more authority even if the lead is read-only. Disable autonomous profiles when you want delegation to remain inspection-only. See Agents and delegation before assigning overlapping implementation work.

Choosing a baseline

Use read-only for local investigation, with delegation disabled or limited to read-only profiles. Use supervised for unfamiliar work and most first runs. Reserve autonomous—and autonomous sub-agents—for bounded tasks in a trusted, recoverable workspace.