Agent Mode

Agent mode is a runtime configuration that shapes how an AI agent operates during a session. It bundles a permission mode — such as prompting for confirmation on risky actions, blocking file writes, or auto-approving everything — with behavioral instructions injected into the system prompt. Think of it as a personality switch: the same agent can be a cautious researcher, an auto-approving editor, or a YOLO-mode wildcard depending on which mode is active.

The mechanism works by loading different instruction sets into the agent's system prompt at startup or mid-session. A plan mode injects language that steers the agent toward read-only research and blocks write operations. An accept-edits mode auto-approves file modifications. A bypass-permissions mode (colloquially YOLO mode) approves all operations without prompting. Modes can be switched during a session, allowing a human to shift the agent's behavior as work progresses.

The tradeoff is trust vs. speed. Permissive modes move faster but risk unintended file changes or destructive operations. Restrictive modes are safer but require more human engagement per step. Teams running agents on production codebases tend to default to plan mode or confirm-on-write; solo developers doing refactoring often prefer accept-edits for throughput.

Open questions include how to formally describe mode semantics in a portable way across different agent runtimes, and whether mode-switching mid-session causes context discontinuities that degrade task coherence. Current practice is largely convention-based rather than standardized.