Roles, models and options

Each agent has options: which model it uses, what role it plays, and whether it asks before it acts. This page explains each option and where to set it.

Figure: Right-click an agent to change its options.

Where to find the options

Right-click an agent. Its menu has these sections:

• Runs: which agent runs in this terminal.
• Agent: Orchestrator and Skip permission prompts.
• Role: No role, the built-in roles, your own roles, and Edit roles….
• Model: Agent default and the agent’s models.
• Account: which login the agent uses. See Usage and limits.

Below them are Configure…, Send prompt… and Restart. Configure… opens the agent sheet, which shows all the options in one place.

Most options apply when the agent starts. If the agent is running, Cuewire tells you to restart it, for example “Restart Builder to apply its new settings”. Right-click it and choose Restart.

Choose a model

Under Model, choose a model, or Agent default to let the agent pick.

The other agents have no list in Cuewire. They use Agent default: choose the model inside the agent itself.

Cuewire can update these lists without a new release. Your organisation can also limit the models you may use. Models it does not allow are not shown.

Roles

A role tells an agent what it is for. Cuewire gives the role’s instructions to the agent when it starts. A role can also have rules, which hold the agent to them whatever it is asked.

Built-in roles:

• A Builder makes the changes it is asked for, in small, tested steps. It leaves other files alone.
• A Reviewer looks for bugs, missing tests and risks, and says what must change. It fixes nothing unless you ask it to.

Give an agent a role: right-click it and choose the role under Role. Choose No role to take the role away.

Note

Claude Code, Codex, Grok and Copilot CLI get a role’s instructions when they start. Gemini CLI, OpenCode and Antigravity take no instructions at launch, so they don’t get them yet. A role’s rules still apply to them where the agent supports them.

Make your own roles

1. Press Ctrl+K (Cmd+K on a Mac) and run Agent roles…. Or right-click an agent and choose Edit roles….
2. Click Add role.
3. Type a Name, up to 40 characters.
4. Under Instructions: what this role does, and what it leaves alone, write what the agent should do.
5. Under Rules, turn on the rules you want (see below).
6. Click Add.

Your role now shows in every agent’s Role list.

Figure: The Agent roles sheet. Built-in roles can’t be changed.

To change a role, click Edit next to it. To delete it, click Remove, then Remove now. Agents that had a removed role keep its name, but not its instructions or rules.

• A changed role reaches a running agent when the agent restarts.
• Your roles follow you to your other computers when settings sync is on.
• In a team, an admin can Share a role with everyone. Shared roles show “From your organisation”. See Teams and organisations.

Role rules: read-only and always asks

A role can hold its agents to two rules:

• Read-only: may read, never change files. The agent starts in its own read-only mode. This works with Claude Code, Codex and Gemini CLI (in its plan mode). Any other agent given a read-only role won’t start, and its node says why.
• Always asks before acting, even with skip permission prompts on. The agent asks you before it acts, even if Skip permission prompts is on for it. Its menu then says Skip permission prompts (its role always asks).

A read-only role also turns off Skip permission prompts for its agents.

Figure: A new role with the read-only rule turned on.

When you review a layer’s changes, Cuewire adds a Reviewer that always runs read-only. See Review your agents’ work.

Skip permission prompts

Agents normally ask you before they run a command or change a file. Skip permission prompts starts the agent with its own option to stop asking:

Warning

With this on, the agent acts without asking you. It can change or delete any file it can reach and run any command. For Codex, it also turns off Codex’s own sandbox. Prefer to run such agents in a sandbox.

Skipping does not apply, or the agent won’t start, when:

• its role is read-only or always asks;
• your organisation allows skipping only inside a sandbox, and the agent is not in one;
• the agent’s own managed settings on this computer forbid it (Claude Code’s managed settings, or Codex’s requirements). The node names the file.

Agents that an orchestrator adds never skip permission prompts.

The agent sheet (Configure…)

Right-click an agent and choose Configure…. The sheet shows:

• Name and Runs
• Role and Model
• Account (for agents that can have one)
• Options: Skip permission prompts and Orchestrator. With skipping on, the sheet warns you and suggests a sandbox.
• Extra arguments
• Session

Save keeps your changes for the next start. Save and restart applies them now.

For the orchestrator, see Agents working together.

Extra arguments

Add options to the agent’s command line, for example --add-dir "my docs". Separate them with spaces. Use quotes around a value with spaces.

• Extra arguments are saved with the canvas. Never put keys or tokens here.
• Cuewire refuses arguments it manages itself, for example --session-id for Claude Code. The sheet says “Extra arguments: argument –session-id is managed by Cuewire”.
• Cuewire refuses commands that don’t start a session you can talk to, for example claude mcp.

Session id and resume

Under Session:

• The current session’s id shows, with Copy to copy it. A new agent shows “No session yet”.
• Resume session by id: paste an id. The agent resumes that conversation at its next start.
• Find a session of this project… lists this project’s past sessions, newest first.

See Pick up a conversation.

What’s next

• Usage and limits
• Sandboxes and safety
• Review your agents’ work