Agents Window

Overview

The Agents window is a separate window built for working with AI agents. While the main editor is where you read and edit code yourself, the Agents window is where you hand work to the assistant and let it run: you start a session against a project, describe what you want, and the agent works through it — reading files, running tools, and proposing changes that you review and apply.

It is designed for running several sessions at once. You can kick off one agent on a project, switch to another, and let them work in parallel, glancing at the list to see which one is running, which needs your input, and which is done.

Using AI sessions needs an active AI entitlement on your account (Basic plan or higher, or an active Trial), the same as the built-in AI Chat. A few features noted below need a higher plan.

Opening the Agents window

• From your project: click Open in Agents in the title bar of the main editor window. This opens the Agents window and carries your current project folder over, so new sessions start against it.
• Command Palette (Ctrl+Shift+P): run Open in Agents (carries the current folder) or Open Agents Window (opens it without a folder).
• Keyboard: Ctrl+Shift+A (in screen-reader mode, Ctrl+Shift+Alt+A).
• Welcome page / banner: the Try out the new Agents window and View All Sessions buttons also open it.

To jump back to a full editor, click Open in AnyAutomation Studio in the top-right of the Agents window. If the current session has a project folder, it opens that folder in an editor window.

The layout

• Title bar. In the center, a clickable session pill shows the active session's icon, title, and folder (and its branch when it uses one). Click it to open the session picker; right-click it for that session's actions. Back and Forward controls (also Alt+Left / Alt+Right) move through recently viewed sessions. Top-right is Open in AnyAutomation Studio.
• Left — the sessions list. Every session, with a New button and search, filter, sort, and grouping controls.
• Center — the active session. The conversation with the agent. A new session starts with the composer, where you pick the project and model and write your first instruction.
• Right — three tabs. Changes (the files the session has modified), Files (a file browser for the session's folder), and Project Explorer (a read-only view of the open TIA Portal project).

Opening any file shows it in a floating editor over the window — that is the standard way files, diffs, and TIA blocks open here.

Starting a session

1. Pick the project. At the top of the composer, click the project chip to open the picker. It lists your recent projects (and recently opened folders); if the list is long, a search box appears. Choose Select… to browse for any folder. Your last choice is remembered for next time.
2. Local or remote. When more than one kind of host is available, the picker shows tabs. Local runs the session on your own machine against a folder on this computer — always available, no setup. Remote runs it on another machine; from the Remote tab you can Connect via SSH (pick a host or type user@host, choose how to authenticate, name it), Connect via Dev Tunnel (sign in, then pick a tunnel), or add a host by address. After it connects, you pick a folder on that machine. Each remote shows a live status, and a Manage entry lets you reconnect, remove, or open its settings.
3. Confirm trust. The first time you target a folder the agent hasn't worked in, you are asked to confirm — the agent will be able to read files, run commands, and make changes there. Declining cancels the selection.
4. Pick a model. The composer's model picker shows the AI model the session will use. It mirrors the model you last chose in the main editor's AI Chat, so both windows stay in sync. See AI Chat for connecting providers and choosing models.
5. Write and send. Type your first instruction in the message box and press Enter (use Shift+Enter for a new line), or click the Send button. The session is created and the agent begins working.

The sessions list

• New session: click New, or press Ctrl+N.
• Find a session: the search icon filters the list as you type; the filter icon lets you sort by created or updated time, group by workspace or by time, show recent or all, and collapse groups.
• Switch sessions: click a row, open the session picker (the title-bar pill, or Ctrl+R) and type to search, or jump with Ctrl+1…Ctrl+8 (and Ctrl+9 for the last one).
• What a row shows: the session's icon and title, its folder, the lines added and removed, and a timestamp — plus its live state:
  • Working — a spinning indicator with a "Working…" line.
  • Input needed — a pulsing dot; when the agent asks permission for an action, an Allow button appears right on the row.
  • Idle / done — a small dot; finished, unread sessions are marked so you can spot them.
  • Failed — an error icon.
• Per-session actions (row hover, right-click, or the title-bar pill): Pin / Unpin, Rename…, Mark as Read / Unread, Mark as Done (archives it) and Restore, Open in New Window, and Copy Worktree Path for isolated sessions. There is no separate delete — Mark as Done archives a session and you can always Restore it.

The message composer

The composer is the box where you tell the agent what to do.

• Writing and sending. Type your request; the box grows with multi-line input. Enter sends, Shift+Enter adds a line, and the round Send arrow (bottom-right) sends once there is text.
• @Codebase auto-context (Pro). With it enabled, the most relevant code from your project attaches to each message automatically as removable chips above the box; remove any you don't want and it won't come back for that message. A small indicator shows whether the index is idle, working, or unavailable. It is off until you turn it on in Codebase Index settings.
• Voice dictation. A microphone button (shown when speech input is available on your machine) lets you dictate by speaking. Click to start, click to stop; your words appear as you talk. If you end a sentence with the configured submit word (by default "submit"), the message sends automatically — you can change or clear that word in settings.
• Attach context. The + button (or Ctrl+/) attaches Files… or an Image from Clipboard. You can also drag and drop files onto the box, paste an image, or type # to reference a file inline. Attachments appear as chips you can click to open or remove.
• Slash commands and skills. Type / at the start of a message for a menu — including /agents, /skills, /instructions, and /hooks, plus your own saved prompts. The Run Skill button (lightbulb) runs one of your skills without typing the command.
• Prompt history. With the cursor at the start of an empty box, press Up to recall earlier prompts.
• Context usage. A Context indicator shows how much of the model's context window the conversation is using; click it for a breakdown.

Repository options

A row of small dropdowns under the composer controls how a session runs against your project:

• Default Approvals — how much the agent may do on its own: Default Approvals (asks before sensitive actions), Bypass Approvals (auto-approves every tool action — shows a warning the first time), or Autopilot (Preview) (the agent works autonomously start to finish; you can stop it anytime). Your organization's policy can disable the elevated options.
• Worktree — whether the session works in an isolated copy of the project or directly in your folder (see Running in isolation below). Available for Git projects.
• Branch — which branch the session starts from; searchable when there are many.

Sub-sessions

Inside an existing session you can start a fresh chat in the same project — useful for a side question or a separate task with clean context. It works like the main composer but skips the project and model pickers, since the workspace is already set.

Running several attempts at once (Best-of-N)

To compare approaches, write your prompt and click Run Best-of-N (the stacked-layers button):

1. Choose how many attempts run in parallel (2 to 5, default 3).
2. Pick a model for each attempt — Same as Current or a specific model per slot.

Each attempt runs on its own private copy of the project so they don't interfere. A Best-of-N bar shows one chip per attempt with its number, model, and live status (Preparing, Running, Done, Error, or Aborted); click a chip to switch to that attempt and read its work. When you've decided, click an attempt's check mark to adopt it — it becomes your active session and the others are discarded. Discard All ends the whole run. After adopting a winner, apply its work with Apply Changes to Parent Repository (below).

Running in isolation

A session can run in an isolated copy of your project so your open files are never touched while the agent works:

• Worktree (default) — the agent works in a separate, throwaway copy; you review its work and merge it in only when you're happy.
• Folder — the agent edits your open project folder directly.

Isolation is available for Git projects, and files belonging to an isolated copy are marked so you can tell them apart.

Reviewing and applying changes

• Changes panel. Toggle it with Show Changes / Hide Changes in the title bar. It lists every file the session created or changed, with added and removed line counts. Click a file to open its diff; Open File and Open Changes are also available per file. When a session is linked to a pull request, an Open Pull Request button appears.
• Change navigation. In the diff header, Previous Change and Next Change step through the changed files, and a counter shows your position (for example, 2 of 7).
• Apply to your project. For a session that ran in isolation, Apply Changes to Parent Repository merges its work into your real project. You'll see Applied changes to parent repository (with an Undo Apply button), or a note that the repository was already up to date, or — if it has diverged — a message asking you to resolve conflicts.
• Undo an apply. Use Undo Apply on the confirmation message, or run Undo Apply to Parent Repository. A dialog confirms the action (it rewinds your project to the earlier point and cannot be undone). For safety, undo is blocked while your project has uncommitted changes or a merge in progress, and is no longer offered once you've committed more work on top.

Approvals and autonomy

While a session runs, the agent asks permission before taking actions — running a tool, editing a file, running a command — and you Allow or deny each request inline (also from the session's row in the list). Certain sensitive operations always ask, even in an auto-approving mode. Set the overall policy per session with Default Approvals in the composer (see Repository options).

Completion sound

When an agent run finishes — or pauses needing your input — while the Agents window is in the background, a short sound plays so you know to come back, with distinct cues for completed, failed, and needs-attention. You can turn this off in settings.

Customizations

A Customizations area in the left sidebar gives the agent extra capabilities and guidance. Each category shows a count and opens a full editor where items are grouped by where they come from (your workspace, your user profile, built-in, or installed plugins) and can be searched. These are the same customizations used by AI Chat, surfaced here for sessions:

• Agents — specialized assistants with their own tools and instructions for specific tasks. Create, edit, enable/disable, or delete.
• Skills — reusable knowledge and workflows for the agent, including built-in ones; a skill of yours can override a built-in of the same name. Some carry a UI Integration tag, meaning a button in the app uses them.
• Instructions — always-on guidance applied across a workspace or your profile, including project rule files the agent reads automatically.
• Hooks — automated actions that run on events in the agent's lifecycle. Configuring hooks needs an open folder.
• MCP Servers — external tool servers that add tools and data sources. Add Server or Browse Marketplace to find them.
• Plugins — bundles of tools, skills, and integrations. Browse Marketplace, Install Plugin from Source, or Create Plugin.

MCP servers and plugins can be turned off by your organization's policy, in which case the panel explains that. See AI Chat for more on these customizations.

Side panels

• Files. A file browser for the session's folder. Click a file to open it in the floating editor; the toolbar offers New File, New Folder, Open Folder (change the root), and Collapse All.
• Project Explorer. A read-only view of the TIA Portal project open in the main window — PLCs, folders, and blocks with their Siemens icons. Click a block to open its source in the floating editor. There are no checkboxes or connect buttons here; to connect to a project, use the main Studio window. When no project is open, the panel points you there.

Notes

• AI sessions require an active AI entitlement (Basic plan or higher, or a Trial); @Codebase auto-context and codebase search require a Pro license or higher.
• Connecting to TIA Portal happens in the main Studio window; the Agents window's Project Explorer is read-only.
• Sessions are archived with Mark as Done, not deleted, and can be restored at any time.