The agent's workspace

Every chat has a workspace: a private computer of its own, with files that last as long as the chat. The agent works there the way an analyst works in a project folder. It brings records in from your connected systems as files, works on them with scripts, SQL and the command line, and keeps its notes in a file. Intermediate results stay on disk instead of in the conversation, so long runs stay on course and every number can be traced to the script and the data that made it.

The workspace is created the first time the agent needs it. You don't set anything up.

Data comes in as files

When the agent needs records from a connected system, it brings them into the workspace as a file: every application with its owner and lifecycle, say, or every open incident on a service. The trace in the chat shows each one as Pulled data, with the file's name and how many rows it holds.

• Each file says where it came from: the system, the query, when it was read, and how many rows it holds.
• Person and email fields are masked on the way in, so names and addresses never reach the workspace.
• Files you add join them. When the agent reads a spreadsheet, document or deck you added to the chat, it copies it into the workspace, where it stays as long as the workspace does. Names and emails in a table are masked the same way, and a document is read page by page. Pictures go to the model directly and never enter the workspace. See Add files to a chat.

The workspace has no internet, and it holds no passwords or keys. It reaches your systems only through Studio, under the same policy as the chat.

Working notes

For longer work, the agent keeps working notes in a file in the workspace: what it set out to do, what it found, and what is left. The notes survive when earlier turns of a long chat are summarised, and when a run resumes after a restart. When the agent says a goal is complete, the goal check reads these notes for the evidence.

Code mode

With Code mode, the agent can write a script that reads your connected systems directly. Reading every application's details, or walking a chain of dependencies, then takes one command instead of hundreds of steps.

• Scripts only read. Any change to a source system still goes through an approval in the chat.
• A script sees only the connections and tools the chat has, and reaches them only while the agent's own command runs.
• Person and email fields come back masked, and every read is in the audit log.

Code mode is on by default. An org admin can turn it off under Settings → Agent Harness → Tools → Workspace. See Agent harness.

Match and relate

Two systems rarely agree on what an application is called. In the workspace, Judge links the records of two files: the same application in your EA repository and in your service catalogue, say. It can also work out a relation no system holds, such as which epics touch which capabilities.

Clear pairs are settled by rules, and Judge decides the rest. The trace shows each run as Judged records. The pairs Judge isn't sure of are left for a person on the Estate tab. Settled matches are kept for your organization, so the next chat doesn't decide them again.

What you get back

When the work is done, the agent publishes what you asked for. It arrives in the chat as a Files ready card, with a download for each file:

• an Excel workbook with its formulas, or a CSV file,
• a PowerPoint deck,
• a Word or PDF document,
• a chart or a diagram as a picture,
• a page: an app built for your question, with its own controls and views.

Open on Stage shows it there, where Download saves the files.

A page is built from the same parts as Studio's visual apps, so it reads like the rest of Studio. The agent draws it in the workspace and looks at the result before it shows you anything. The page opens on the Stage, works there with no network, and says where its data came from and when it was read.

The Estate tab

The Estate tab on the Stage shows the data the agent brought in from your systems. It appears once the chat holds some, with a count of what waits for you.

• Data: each file's rows, and each record with its neighbours.
• Graph: how the files relate to each other.
• Review: the pairs Judge left for a person, such as "Same record?" or "Related?". Accept or Reject each one. Your decision is kept for your organization, and later chats use it. Pairs not judged yet can go to Judge with Ask Judge.

Files the agent wrote itself, and files you added, aren't listed here.

How long it's kept

A chat's workspace is deleted 7 days after its last use, and when the chat is deleted. What it published stays in the chat. If you come back to an older chat, the agent starts a new workspace and brings in what it needs again. A forked chat starts a workspace of its own too.

Sub-agents work in the same workspace as the chat that started them.

Workbenches have one too

Each workbench has its own workspace, kept for 30 days after its last use. The job's folders are real files there, each stage reads only the inputs its brief declares, and data a stage brings in is there for the stages after it.

Limits

Each organization has 50 sandbox hours a month, counted while a workspace is running. The Workspace row under Settings → Agent Harness → Tools shows this month's use. Scripts in Code mode can make up to 20,000 calls to your systems per organization per day. See Billing, trial and limits.

See also: Co-create with agents · Agent harness · Security & data.