Simulator & Mac

iOS Simulator driving (Pro)

• Boots simulators, builds and installs your app, taps/types/swipes through the real UI tree, reads logs, and screenshots into the Artifacts pane.

Mac computer use (Pro)

• Drives GUI-only Mac apps (Notes, Mail, System Settings…) natively via Accessibility, with a per-conversation app allowlist and per-action approval. You do not have to allow an app in advance: the first time the agent targets one that is not on the list, Naseem asks you once in approval mode, and adds it by itself in Autonomous mode.
• What it can do — read a window as a compact tree of named elements, click and double-click, right-click for a context menu, type into fields (appending or replacing), send shortcuts and key sequences, scroll, and wait for an element to appear. It reads the result back after typing, so a "typed" result means the text is really there.
• Scripting or clicking, by the job — for a scriptable app doing data work (TextEdit, Finder, Mail, Notes, Numbers) one AppleScript is faster than clicking, and the agent uses it. For apps with no scripting dictionary, and whenever the window state matters or typing must be verified, it drives the app through Accessibility. Measured on the same TextEdit task: one script call in half the time, but with AppleScript's carriage returns in the file; the Accessibility path took more calls and produced clean text.
• Several steps in one call — when the agent already knows the next few actions (open a menu, pick an item, type, press Return, wait for the result) it sends them as one batch. Each step is verified, the batch stops at the first failure and says which steps did not run, and the window state comes back with the result. Fewer model round trips is what makes a GUI task fast.
• What it does not do by default — it does not look at pixels. The agent works from the same accessibility tree that VoiceOver uses, which is why it needs the Accessibility permission and why apps with poor accessibility (some Electron and cross-platform apps) are harder to drive. For those, switch on Allow screenshots of Mac apps in Settings → Integrations: the agent can then capture an app's front window when the tree is empty or when appearance is the question. Needs Screen Recording permission and a vision-capable model, and each screenshot costs image tokens.

Board (Pro · Beta)

• Up to four conversations side by side. New in 1.4.23 and marked Beta: it works end to end and is measured, and the details will keep improving with your feedback. Click the grid button at the left of the toolbar to enter the Board. Each cell is a full conversation: its own transcript, composer, model, profile and autonomy controls in the cell's header. Running or idle, any conversation qualifies — compare two models on the same prompt, or watch four long runs at once.
• Adding and removing. Right-click a conversation → Add to Board, or use the + menu in the Board's toolbar. The layout follows the count: one fills the area, two sit side by side, three make a top pair and a full-width bottom, four make the grid. Rows on the Board carry a small grid badge. Remove from the cell's ⋯ menu or the row's context menu.
• The active cell has the accent border and is selected in the sidebar; click anywhere in a cell to make it active, or click its row in the sidebar. ⌘1 to ⌘4 jump between cells in reading order. Return sends there, ⌘F searches there, and the Files | Artifacts pane on the right shows that cell's project files and artifacts. Drag the dividers between cells to resize them. Double-click a cell's title to open that conversation on its own; clicking a sidebar row also leaves the Board. The Board and its members are remembered across relaunches.

Sub-agents & Telegram (Pro)

• Sub-agent delegation — the agent spawns focused child agents for big side tasks and works in parallel. The transcript says so plainly: a sub-agents working line with a spinner while they run, and one card per sub-agent you can expand for its steps.
• Models by phase — plan and verify on the strong model, execute on the cheap one. In a custom profile (Profiles → Edit) pick a sub-agent model — child agents run on it; the default is the helper model, a fast sibling of your chat model — and optionally a review model, a judge that checks the final answer against the tool trail before the agent may say it is done, at most twice per run. Both are Pro.
• Telegram remote — talk to your agent from your phone, including remote Approve/Deny of gated tool calls (Settings → Integrations).