Coding harnesses

AgentGrid treats each supported AI CLI as a harness. A harness can run as a full interactive pane, as a worker spawned by a master, or both. All harness panes use the same project folder and live alongside your other development tools.

Harness catalog

Master MCP indicates that a pane can coordinate workers and use AgentGrid's canvas, terminal, and browser tools. A harness without master MCP can still run interactively and can be chosen as a worker by another master.

Add and arrange harnesses

Open Settings -> Agents to enable, disable, and reorder harnesses. The first three enabled harnesses appear directly in the spawn menu with Cmd+1Ctrl+1 through Cmd+3Ctrl+3. Additional harnesses appear under Other agents.

AgentGrid detects whether each CLI is installed. Use the in-app Install and Sign in actions when a harness is not ready; you can complete both without leaving AgentGrid. See Coding harness setup for troubleshooting.

What runs in a pane

An interactive harness pane runs the provider's CLI with your environment and the active space's project folder as its working directory. AgentGrid provides the surrounding pane controls, transcript handling, search, paste behavior, and orchestration wiring.

Claude sessions can be reopened from Sessions -> Resume Claude in the spawn menu. Recoverable worker sessions across installed harnesses are available to a master through list_recoverable_sessions and resume_worker.

Mix harnesses and models in one team

A master can start a worker on any installed harness, even when it differs from the master's own runtime. Pass harness to spawn_worker or spawn_role, and optionally pass a harness-specific model.

spawn_role({
  role: "builder",
  harness: "claude",
  model: "claude-sonnet-4-6",
  prompt: "Implement the settings change."
})

spawn_role({
  role: "qa",
  harness: "codex",
  model: "gpt-5.6-sol",
  prompt: "Review the builder's diff."
})

Omit harness to use the master's harness. Omit model to use the harness's configured default. A master can call list_worker_harnesses for the installed catalog and current model guidance.

Permission modes

Settings -> Agents also controls the default permission mode for new agent work:

• Default uses each harness's normal behavior.
• Auto accepts safe tool calls after checking them for risky actions and prompt injection, and blocks calls that fail that review.
• Plan asks the agent to propose a plan before making changes.
• Bypass Permissions disables permission checks and is intended for controlled automation environments.

Choose the least-permissive mode that still supports your workflow.

Harness-specific features

The shared pane surface does not mean every provider exposes identical capabilities. Voice calls and direct clipboard-image attachment are capability-gated, and session-resume behavior follows what each underlying CLI supports. AgentGrid only shows actions that apply to the selected harness.

Related

• Terminals - terminal input, search, paste, and scrolling behavior.
• Agents and workers - the master/worker mental model.
• Orchestrating agents - role-based parallel work and review loops.
• Voice mode - live voice conversations on supported panes.