Chat & Sandbox

Overview

The Agent tab is the primary workspace — where you talk to your agent and watch it work. It features a split-panel layout with the chat on one side and a live desktop preview on the other.

This is where the magic happens: you type a message, and the agent takes action on a real computer in real time.

The Split-Panel Layout

The Agent tab shows a side-by-side layout with the chat interface on the left and the live desktop preview on the right.

You can resize the panels by dragging the divider between them to find the balance that works for you.

Collapsing the Desktop Preview

If you prefer a full-width chat experience, you can hide the desktop preview entirely:

• Drag the divider all the way to the right to collapse the preview panel
• Keyboard shortcut — Press ⌘ \ (Mac) or Ctrl \ (Windows) to toggle the preview on/off
• Click the handle arrow on the divider to expand/collapse

Your preference is remembered between sessions — if you collapse the preview, it stays collapsed next time you visit the agent.

This is useful when you want to focus on the conversation without the visual distraction of the live desktop, or when you're working on a smaller screen.

The Chat Interface

Type your message in the input area at the bottom and press Enter to send. You can ask the agent to do anything a human could do at a computer:

• "Go to Google and search for the latest AI news"
• "Write a Python script that analyzes this CSV file"
• "Log into our CRM and export last month's leads"
• "Install ffmpeg and convert that video to MP4"

The chat displays your messages, agent responses, tool calls (actions the agent is performing), screenshots of the agent's screen, and web search citations.

Keyboard Shortcuts

The Sandbox Preview

The sandbox preview shows a live stream of the agent's desktop — a real Linux environment with a window manager, browser, terminal, and applications. You'll see the agent's mouse cursor moving, windows opening, browser tabs loading, and terminal commands executing in real time.

When waking an agent, you choose the screen resolution for the sandbox. Higher resolutions provide more screen real estate for the agent to work with but use more bandwidth for the live preview.

Waking and Sleeping

Wake Agent

Click the Wake Agent button to provision a sandbox. The process takes a few seconds — a sandbox is provisioned in the cloud, the Linux desktop boots up, and the live preview connects.

Sleep Agent

Click Sleep to shut down the sandbox. A confirmation dialog appears because session files that haven't been synced to persistent storage will be lost, and any active processes will stop.

Sandbox Persistence (Persistent by Default)

By default, sleeping pauses the sandbox — all state is frozen and preserved:

• Sleep becomes Pause — The confirmation dialog reassures you that all state is preserved instead of warning about data loss
• Everything is frozen — Files, processes, browser sessions (including website logins and cookies), installed packages, and memory are all preserved
• Resume in ~1 second — Click "Resume Agent" to reconnect to the frozen sandbox instantly, instead of provisioning a fresh one
• No compute costs while paused — The sandbox doesn't use resources in its paused state

This means agents that maintain browser sessions (logged-in websites, open tabs), have heavy package installations, or run background services keep everything intact across pauses. Instead of spending the first few minutes of every session logging in and reinstalling tools, the agent picks up exactly where it left off.

If you prefer a fresh environment each time, you can switch to On-Demand mode in Settings → Sandbox. In on-demand mode, sleeping destroys the sandbox entirely and a new one is provisioned on the next wake.

Onboarding Flow

New agents start with an onboarding conversation. The agent introduces itself and walks you through setup conversationally — asking about its purpose, communication style, schedule, and email preferences. It configures each setting as you discuss it using settings_manager, and sends a summary email when complete.

During onboarding, only settings_manager and send_email are available. This ensures the agent focuses on setup before taking action.

The Action Queue

While the agent is working, you'll see its actions in the chat as expandable tool-call blocks. These show exactly what the agent is doing — screen interactions, terminal commands, web searches, and data extraction. Each action block is collapsible so you can focus on the results without the noise.

Actions can be saved, recorded, and replayed as deterministic sequences. For full details on building and managing action sequences, see Actions & Sequences.

Beyond the Dashboard

The dashboard chat isn't the only way to interact with your agent. You can connect external messaging platforms via the Channels tab — starting with Telegram.

When connected, you send messages directly in Telegram and receive responses from your agent in real time. Your agent works the same way as in the dashboard — same tools, skills, and context. If the agent is sleeping, a Telegram message automatically wakes it up — no need to visit the dashboard first.

This is useful for mobile-first workflows, quick interactions on the go, or any situation where opening the dashboard isn't practical.

Tips & Best Practices

• Be specific — Instead of "research companies", say "Go to Crunchbase, find the top 5 AI startups founded in 2025, and save them as a dataset"
• Use the live preview — Watch the agent work to catch issues early and give real-time corrections
• Collapse when chatting — Hide the preview for a clean, focused chat experience; expand it when you need to see the agent's screen
• Let the agent install tools — If it needs ffmpeg, Python packages, or browser extensions, it will install them automatically
• Don't micromanage — Give the goal, not every step. The agent figures out the best approach

What's Next?

• Actions & Sequences — Record, build, and replay deterministic action sequences
• Channels — Connect Telegram so you can chat with your agent from external platforms
• Context & Configuration — Configure skills, instructions, and tools to shape your agent's behavior
• Files — Understand session files vs. persistent storage
• Credentials — Securely store passwords and API keys