Skip to content
Titan Docs Titan Docs

What the agent can see

Overview

Section titled “Overview”

Most surprises with the workspace agent come from a mismatch between what you think it can see and what it actually reads. This page is the canonical answer. If you are about to use the AI panel for the first time, read this first.

For the AI Chat surface, menus, and limits, see Workspace agent.

What it can see

Section titled “What it can see”

When you send a message in the AI Chat panel, the agent receives:

  • The path of the file you currently have focused in the editor area. This is how it knows which file you mean when you say “this file.”
  • The full file tree of the current project. Names, paths, and folder structure of every file. Not the contents.
  • The contents of files you explicitly attach to a message using the attachment control in the prompt input. Attachments are scoped to the conversation, not persisted across page refresh.
  • The contents of files the agent itself opens via tool calls during the conversation, if the model supports tool use.
  • Prior messages in the current conversation. A new conversation starts with no memory of older ones, even in the same project.

What it does not see

Section titled “What it does not see”
  • Files in other projects. The agent’s view is scoped to the project you have open.
  • Anything outside Titan. Your browser tabs, your file system beyond the project folder, your other apps. Closest analogue: a Cursor session has the same boundary.
  • Comments threads on documents. They live in the comments tab, not in the agent’s context. If you want the agent to act on a comment, paste the relevant text into the prompt.
  • Version history. The agent does not read past versions of files. It reads the current state.
  • Other users’ selections or cursors in a shared project.
  • The Kanban board’s tasks as automatic context. TODO(verify): board content as agent context is on the roadmap; until shipped, paste task text into the prompt.

Why this matters

Section titled “Why this matters”

If you ask “summarize everything in this project,” the agent will only summarize files it can actually read in this turn. For a project with many files, that usually means the file tree plus the files you have open or attached. It is not a search across every file’s contents by default.

The practical move is to attach the specific files you want the agent to use. Three small attachments give a better answer than asking the agent to “look at the project.”

Choosing a model

Section titled “Choosing a model”

The model selector in the AI Chat input picks which provider answers. Some providers run in chat-only mode and cannot perform file actions (open a file, propose an edit). If you ask for an action and get a description of the action instead, check the model — see Models and AI settings and Agent actions and approvals.

Privacy implications

Section titled “Privacy implications”

The agent sends what it can see to the model provider you have selected. Browser projects do not leave your device for storage, but if you send a prompt to a cloud model, the prompt and its attachments go to that provider. If you want the conversation to stay on-device, run a local model. See Models and AI settings.

Section titled “Related Goals”