Custom Agents

Mobile Agent ships with three native agents — General, Plan, and Build — but it's built for customization. You can create your own agents with a custom system prompt, choose whether they can act as subagents, and control exactly which tools and MCP servers each agent may call.

Native agents

• General — the balanced default for everyday questions and tasks.
• Plan — analyzes first, produces a plan, then executes it step by step. Great for multi-step work.
• Build — mode-focused implementation agent for writing, editing, and iterating on files.

Native agents can also be granted subagent visibility: a primary agent may delegate sub-tasks to a subagent when the subagent flag is enabled on it.

Create an agent

1. Open Settings → Agents → New agent.
2. Give it a name, description, and the system prompt that defines its behavior (tone, responsibilities, constraints).
3. Optionally attach docs — reference files the agent carries as context.
4. Configure its model (provider + model) and temperature.
5. Set the mode:
   • primary — selectable as the main agent for a conversation.
   • subagent — only callable as a delegate from another agent.
   • all — usable in both roles.
6. Toggle hidden to keep it out of the picker, and enabled to allow use.
7. Save. The agent now appears in the composer's agent picker when its mode permits.

Per-agent tool permissions

Each agent has its own permission matrix (toolPermissions):

• Built-in tools — enable/disable individual tools (workspace reads, folder writes, download, todos, schedules, question, …) per agent see the Built-in Tools reference.
• MCP servers — control which connected MCP servers this agent may use.

A sensitive workflow agent, for example, can be locked to read-only tools and a single trusted server, while a build agent gets full workspace write access.

Custom system prompts and docs

The prompt field is injected into the model as the agent's system prompt. Beyond that, agents can carry docs: attached reference files that provide detailed context (style guides, API references, permission rules) without bloating the one-line prompt. Multiple agents can share docs, and each doc is versioned with created/updated timestamps.

Import and export

Agents are stored as markdown:

• Export produces a .md file with the agent's prompt, docs, and configuration in frontmatter.
• Import restores it on another device or account.

This makes agents easy to share, back up, and version in a repository alongside your skills.

Start a conversation per agent

From the composer's agent picker (or an agent's detail screen), select the agent and start a chat. The conversation inherits the agent's model, system prompt, and tool permissions for the duration of the run.

Tip: Combine custom agents with the Scheduler — schedule a job against a custom agent so background runs also respect its permissions and prompt.