--- title: "Agent Mode" canonical: https://trickest.com/docs/using-the-app/ai-agent/agent-mode description: "Open the full-screen agent workspace, start a chat, connect a workflow, and work beside the canvas." --- # Agent Mode ## Overview **Agent Mode** is the full-screen workspace for chatting with the Trickest agent. From one shell you start conversations, attach a workflow, and watch the agent plan, edit the graph, and run jobs. The canvas and related surfaces sit in a workbench next to the chat. This is not the **AI Agent** node you place on a workflow canvas. That node is a step inside a DAG. Agent Mode is the chat workspace you open from the platform sidebar. ## Open Agent Mode Sign in and stay on any platform page that shows the main sidebar. Click the **Agent Mode** card at the bottom of the sidebar. The description reads **Chat and automate**. You can also open [trickest.io/agent](https://trickest.io/agent). To leave Agent Mode, click **Back to Platform** at the bottom of the navigator. That returns you to the rest of the app. The Agent Mode card appears when your workspace can use the in-app agent. If you cannot open it, ask a workspace admin or check [Bring Your Own Key](/docs/using-the-app/ai-agent/bring-your-own-key). ## Understand the layout Agent Mode uses three panes on a wide screen: 1. **Navigator (left).** Session list, search, and the **Agents**, **Tasks**, and **Skills** destinations. 2. **Chat (middle).** One conversation at a time. Type here, attach files, pick a model, and connect a workflow. 3. **Workbench (right).** Surfaces for the connected workflow, including Editor, Database, Files, and Terminal. On a narrow viewport the navigator and workbench open as sheets. The chat stays on screen. ### Navigator | Control | What it does | |---|---| | **Agents** | Open the new-conversation composer or return to chat. | | **Tasks** | Create and manage scheduled agent runs. | | **Skills** | Create or import skills the agent can load during a chat. | | **Sessions** | List recent conversations, or group them by workflow. | | **New Agent** (`+`) | Start a new conversation. | Right-click a session (or open its menu) to **Rename**, **Pin**, or **Delete** it. ### Keyboard shortcuts These shortcuts apply when you are not typing in an input, editor, or dialog: | Shortcut | Action | |---|---| | **Cmd+B** (Mac) or **Ctrl+B** (Windows/Linux) | Show or hide the navigator | | **Cmd+J** or **Ctrl+J** | Show or hide the workbench | | `Cmd+\` or `Ctrl+\` | Reset pane widths | | **Cmd+K** or **Ctrl+K** | Focus session search | ## Start a conversation Click **Agents**, or the **+** next to **Sessions**. Use the **Workflow** control in the composer to pick an existing workflow, or leave it on **No workflow** for a chat that is not bound to a workflow yet. Open the model picker in the composer and select a model your workspace can use. Describe the goal. Press **Enter** to send, or **Shift+Enter** for a new line. Use **+** to attach files or images. You can start from a library template instead of a blank graph. Under **Start from a workflow**, pick a published workflow. Trickest copies it into your Playground workspace when that workspace exists; otherwise it asks you to choose a workspace. The copy attaches to the chat and a starter prompt appears in the composer. Sending is up to you; the copy does not run on its own. Recent sessions appear under the composer so you can resume work without hunting in the navigator. Connecting a workflow before the first send opens the workbench so you can inspect the canvas while you draft. ## Connect a workflow A conversation can stay unbound, or it can attach to one workflow at a time. Binding scopes the agent's work and the workbench to that graph. **From a new chat.** Use the **Workflow** picker in the composer. Search by name, project, or workspace. Rows group by workspace when you have more than one. **During a conversation.** The same picker stays in the composer. You can connect, switch, or choose **No workflow** to disconnect. Connecting a workflow: - The workbench opens on that workflow's **Editor** tab. - If the workflow lives in another workspace, Agent Mode switches to that workspace. - The agent can inspect nodes, edit the graph, and run jobs on that workflow. - Selecting a node on the canvas offers an **@** mention so you can refer to it in the next message. Disconnecting hides the workbench. Saved tabs for that workflow stay available the next time you attach it. The agent can also create a new workflow or switch the conversation onto a different one as it works. The picker and the session list update to match. ## Work in the workbench The workbench appears after you connect a workflow. Launchers at the top open these surfaces: | Tab | Use it to | |---|---| | **Editor** | See and edit the workflow canvas. Nodes appear here as the agent builds. | | **Database** | Browse Live Tables produced by the workflow. | | **Query** | Run TQL against those tables (shown when the workflow has live data). | | **Files** | Browse sandbox and workflow files, including plan documents. | | **Preview** | Open an app or report the agent is serving. | | **Terminal** | Use a shell in the workflow sandbox, or follow agent activity. | An empty Editor tab points you back to chat: describe the goal there, and the agent fills in nodes and connections. You can still add a node from the canvas if you want to drive a step yourself. Layout controls sit next to the workbench tabs: expand the workbench, restore the split, or hide it to focus on chat. Hiding the pane does not close its tabs. The Editor tab is the canvas inside Agent Mode. For the full workflow editor (runs history, execute controls, and the rest of the platform chrome), use **Back to Platform** and open the workflow from there. See [Navigating the Editor](/docs/using-the-app/workflow-and-executions/navigating-the-editor). For how the agent's Linux environment, long-running commands, and previews behave, see [Agent sandboxes](/docs/using-the-app/ai-agent/sandboxes). ## Guide the agent while it runs While a reply is in progress: - **Enter** queues a follow-up for the next turn instead of starting a second run. - You can steer the current turn with additional instructions. - Stop ends the in-flight run. For larger work the agent writes a plan file and pauses for a decision. **Build** approves it and the agent executes the steps. **Reject** sends your feedback so it can revise. Those controls sit on the plan card in chat and on the plan file in the workbench, so you can still decide after hiding the workbench. You can open referenced nodes, files, and tables from the transcript. Those links reveal the matching workbench tab. ## Tasks and Skills **Tasks** in the navigator schedule agent runs: when they fire, which model they use, and the instructions they send. They are separate from the to-do list the agent keeps during a live chat. **Skills** are reusable instructions you create, import, or edit under `/agent/skills`. The agent can load a skill during a conversation. Opening Skills keeps the navigator visible and uses the main pane for the list and editor. ## Related Linux workspace, files, commands, and app previews the agent uses. Connect a provider key when your plan bills the agent to your account. Canvas layout and execution controls in the full workflow editor. Copy a published workflow into a workspace from the Library page. --- _Markdown source of https://trickest.com/docs/using-the-app/ai-agent/agent-mode._