Your first conversation

Start with a small task whose result you can inspect. The goal is to learn what your selected model, enabled tools, and permissions can do together—not to enable every feature before the first message.

Check the basics

Complete setup on Home, then choose Start chatting, or open a new conversation. Check the selected model and permission mode before sending anything.

For initial testing, use Ask. It requests approval before sensitive actions. FULL+ permits task-relevant system actions without those additional ABot prompts; do not select it merely to make a test finish faster. An existing conversation retains its own saved mode, so check the conversation you are actually using.

Try a request with a visible result

For a first connection check, no tool is necessary:

Explain the difference between a language model and an agent runtime
in five short sentences. Do not use tools.

A response confirms that this request reached the model and returned through ABot. It does not establish that file access, desktop access, memory, or every tool works.

Next, try a small research task with an explicit source and output. Replace the URL below with a public page you want to read:

Read https://example.com/ and summarize what the page says.
Give me three bullet points and include the source URL.
If fetching fails, explain the failure instead of guessing.
Do not create or edit any files.

This task requires the enabled Web plugin’s web_fetch tool. Fetching a known public page does not require a search API key. Searching for pages is a separate operation with different coverage and configuration.

Read the activity, not just the answer

Follow the conversation’s tool activity and results. Check whether ABot actually fetched the page, whether the tool succeeded, and whether the final response is supported by the returned content.

The same principle applies to files and commands: a confident sentence is not evidence that an artifact was written or an application opened. Inspect the saved file or requested outcome when the task matters. A failed tool can still be followed by a useful explanation, but the explanation should not be mistaken for successful execution.

If the model chooses the wrong action, narrow the request and retry. A clear boundary such as “read only” helps communicate intent; the Runtime’s permission controls remain the actual approval boundary for sensitive actions.

Add a working folder when you need one

Use a project when several conversations belong to the same folder. Open Chats → New project. Choose an accessible folder, optionally name the project, then choose Create project. Browse lists directories on the Runtime computer, which can differ from the computer running your browser.

For a first file task, use a disposable working folder with non-sensitive content. Ask for one named result, then open it yourself to confirm the content.

Read notes.md in this project's folder.
Summarize its open questions in a new file named questions.md.
Leave notes.md unchanged. If questions.md already exists, ask me
before replacing it.

Only try this after creating notes.md and enabling the relevant file tools. A project folder organizes working context; it is not an operating-system sandbox for approved shell commands. See permissions and projects before granting broader access.

Steer, stop, or continue

During an active request, a text-only draft uses Steer current request as its primary action. For a separate follow-up, choose More send options → Send next. A draft with attachments uses Send next rather than steering. Use Stop response to stop ongoing work; stopping is not an undo button for actions that already completed. The Web UI walkthrough explains these controls and their checkpoints in detail.

Keep a follow-up in the same conversation when it relies on the previous result. Start a new conversation for unrelated work. Optional memory can carry selected information across conversations, but it is not required to continue a normal conversation.