# Permissions and approvals Source: https://docs.ideharmony.com/security/permissions An approval prompt is the last point at which a bad instruction is cheap to stop. Everything else on this page follows from that. ## Where instructions come from An agent's next action can be influenced by any of these, and none of them are trustworthy: | Source | Why it is untrusted | |---|---| | Model output | Can be wrong, or steered | | Repository content | Written by whoever wrote the repo | | Web pages in a browser surface | Arbitrary | | Provider output | Passes through systems you do not control | | Extension and skill code | Runs with real privileges | | Messages from peer machines | Originated elsewhere | > [!SECURITY] > Treat all of the above as attacker-controlled input. An approval gate is not > bureaucracy; it is the boundary between "something suggested this" and "this > happened to my machine." ## Modes Execution mode governs how much an agent may do unprompted: ```text /mode safe /mode ask /mode yolo ``` | Mode | Behaviour | |---|---| | `safe` | Conservative; consequential actions gated | | `ask` | Prompts before acting | | `yolo` | Minimal friction | ## YOLO defaults to on > [!WARNING] > The desktop's YOLO setting defaults to **on**, and it injects provider > full-access flags at launch. That removes friction, and it removes a safety > rail at the same time. Before pointing agents at a repository you care about, > look at **Settings → Agents / Harmony** and decide deliberately. YOLO is an authority setting. It is separate from Crew Mode ([Experimental](/docs/reference/feature-status)), which is about delegation scale. Turning one on does not turn the other on. ## A2A mode Agent-to-agent coordination has its own three-way control: | Mode | Meaning | |---|---| | `Off` | Agents do not message each other | | `Connect` | Agents may message each other | | `Orchestrate` | Agents may coordinate work | Loop prevention watches conversations, but the mode is the real control. If you do not want agents talking, `Off` is more reliable than hoping they stay on topic. ## Reading an approval Three questions: 1. **What exactly is being asked for?** A path, a command, a network call. 2. **Does it match what I asked the agent to do?** Scope creep shows up here first. 3. **Is it reversible?** Reversible actions can be approved faster. > [!IMPORTANT] > Approving quickly, repeatedly, trains you to stop reading. If you find > yourself clearing prompts without looking, that is a signal to change mode > rather than to keep going. ## Permission ceilings for extensions An extension cannot grant itself more authority than its ceiling permits. This is what keeps a read-only skill from acquiring shell access. It is not a sandbox — see [security overview](/docs/security). ## Remote messages are contained Messages arriving from peer machines over LAN Link render as **text only**. They are never pasted into a PTY and never routed to an execute path. A remote peer cannot make your machine run something by messaging you. ## Related - [Security overview](/docs/security) - [Permissions and approvals concept](/docs/concepts/permissions-and-approvals) - [Customization](/docs/customization)