# AI Agents





## Overview [#overview]

Run AI coding agents in isolated workspaces. Each agent works independently without interfering with others.

<img alt="AI Agents" src="__img0" />

## Supported Agents [#supported-agents]

* **Claude Code** - Anthropic's CLI
* **Amp** - Amp Code CLI
* **Codex** - OpenAI's assistant
* **Gemini CLI** - Google's CLI agent
* **Mastracode** - Mastra's terminal coding agent
* **OpenCode** - Open-source alternative
* **Pi** - Minimal terminal coding harness
* **Copilot** - GitHub's coding agent
* **Mistral Vibe** - Mistral's terminal coding agent
* **Kimi Code** - Moonshot AI's CLI coding agent
* **Grok** - xAI's CLI coding agent (Grok Build)
* **Cursor Agent** - Cursor AI agent
* **Droid** - Factory's autonomous coding agent
* **Polygraph** - Meta-harness with cross-repo visibility and persistent memory
* **Hermes** - Nous Research's autonomous coding and research agent
* **Muse Code** - Meta's terminal coding agent with persistent subagents
* **Devin** - Cognition's Devin agent in the terminal, with Devin Cloud handoff
* **fx** - Vercel's terminal coding agent
* **Antigravity** - Google's Antigravity CLI (`agy`)
* **Kiro** - AWS's spec-driven CLI coding agent

## Chat [#chat]

Superset includes a built-in chat panel for AI interactions without leaving the app.

* Use the model picker to choose supported models and connect providers
* Chat session titles are generated automatically from the conversation
* Use the thinking level control to adjust reasoning effort for the selected model

## Models and Reasoning Effort [#models-and-reasoning-effort]

The model pickers include **GPT-6 Sol** and **GPT-6 Luna** for Codex and
Superset Chat, and **Opus 5.5** for Claude Code and Superset Chat. Available
reasoning levels depend on the agent and model; see the [launch options](/cli/cli-reference#agent-effort-levels).

## Agent Settings [#agent-settings]

Go to **Settings → Agents** to customize launch behavior for each built-in agent.

* **Enabled** - Show or hide the agent in launchers
* **Command (No Prompt)** - Base command for terminal agents
* **Command (With Prompt)** - Prompt-aware command for terminal agents
* **Prompt Command Suffix** - Optional flags appended after the prompt payload
* **Task Prompt Template** - Template used when launching an agent from a task
* **Resume Args** - Arguments used to restore a killed session by its id (e.g. `--resume`)
* **Model Override** - Optional model override for Superset Chat
* **Reset to Defaults** - Restore the shipped configuration for that agent

## Usage [#usage]

Run agents directly in the terminal:

```bash
claude "Add error handling to fetchUser"
```

## Auto-Run Command [#auto-run-command]

When launching an agent from a task, the **Auto-run command** toggle controls whether it executes immediately.

* **On** - Terminal agents run their command right away, and Superset Chat sends the task prompt immediately
* **Off** - Terminal agents stage the command without pressing Enter, and Superset Chat opens as a draft instead

The setting is remembered between sessions.

## Resume Interrupted Sessions [#resume-interrupted-sessions]

If an agent's terminal dies under it (a crash, a reboot, a killed process), Superset resumes the session automatically: the pane relaunches the agent with its resume command (e.g. `claude --resume <session-id>`) and shows a "Resuming…" pill while it happens. Only unexpected deaths qualify: sessions you close or kill deliberately, and agents that exited cleanly, stay down.

If the resume fails, the pill turns into "Failed to resume" with a **Retry** button and a dismiss control.

Works for agents with an id-based resume command (the **Resume Args** field under **Settings → Agents**). From the CLI:

```bash
superset agents create --workspace ws_… --agent claude --resume-session <sessionId>
```

## Parallel Agents [#parallel-agents]

Create multiple workspaces and run an agent in each. Compare results and merge the best solution.

## Inline Prompts [#inline-prompts]

Agents can pause and ask questions, request tool approval, and ask for plan review directly in the chat pane.

## Sandbox Access [#sandbox-access]

When an agent needs access to files outside its workspace, it will prompt for permission in chat so you can approve or deny the request inline.

## Notifications [#notifications]

Get notified when agents finish. Configure sounds in Settings.

## Agent Skills [#agent-skills]

Teach any agent to drive Superset itself with the official [skills](https://github.com/superset-sh/skills), one for the `superset` CLI and one for the [MCP server](/mcp-server):

```bash
npx skills add superset-sh/skills
```