DocsAgentsQuickstart

Your first agent run.

From a blank workspace to a finished deliverable in roughly five minutes. Create a project, upload a file, ask the agent to do something useful, and watch it work.

§ 01Create a project

Open the agents workspace and click New project. A project is the top-level container — it holds your files, your runs, and the agent's history. Give it a name that reflects the work, not the model: "Acme Q3 contracts" is better than "test".

You can have many projects. Each one is fully isolated from the others — uploaded files, terminal scrollback, browser cookies, and run history never cross between projects, even when the same agent type is in use on both sides.

TipPick the project shape early. A project per account, per deal, or per matter is usually the right granularity. A project per question is too small.

§ 02Upload files to the drive

Every project has a drive — its own file area. Drag files in from your desktop or use the upload button. Common formats are supported out of the box.

Each upload is validated by magic-number signature, not just extension, and the original filename is retained as metadata. Once a file is in the drive the agent can read it as soon as you reference it in chat.

§ 03Start a chat with the agent

Click into the chat pane. Ask the agent something concrete — concrete enough that you'd know whether it succeeded.

prompt
Read the contract I just uploaded. Pull out the parties,
effective date, term length, renewal terms, termination
clauses, and any payment schedules. Give me a one-page
summary I can paste into a memo.

Hit send. The agent receives your message together with a context bundle: the project's file list, the active surface, and any relevant skills. It plans, then begins to act.

§ 04Watch the surfaces switch

This is the moment that distinguishes the agent workspace from a chat product. As the agent works, the UI follows along.

You don't have to switch tabs or click anything. The agent emits a surface_switch event each time it changes context, and the UI follows.

§ 05Inspect artifacts

When the agent finishes, two things have happened.

Artifacts are the part of the run that survives the chat. The chat is the conversation; artifacts are the deliverables.

§ 06Re-run from history

Open the history panel for the project and you'll see every run, with timestamps and the prompt that started each one. Click into a past run and you can:

NoteHistory is a permanent audit trail. Even when you re-run, the original run is preserved — you get a new run record, not an in-place mutation.

§ 07Common patterns

A few requests that almost always produce a clean result on the first run.

Summarize this contract

Drop a contract PDF into the drive. Ask the agent to extract parties, effective date, term, renewal, termination, and payment terms, and to produce a one-page memo. The agent reads the document, identifies the clauses, and writes a structured summary.

Extract invoice line items

Drop a folder of invoice PDFs into the drive. Ask the agent to extract vendor, invoice number, date, line items, subtotal, tax, and total from each one, and to produce a CSV. The agent runs through the files, OCRs as needed, and saves the spreadsheet as an artifact.

Research and brief

Ask the agent to research a company, a market, or a topic. It will use the browser surface to gather sources, the terminal to process them if useful, and produce a written brief with citations.

§ 08Where to go next

Three good directions from here.