Start a thread

A thread is where you work with an agent. You ask a question or give a task, and the agent answers, calling its tools and skills as needed to work with your data, files, workflows and canvases. People write in threads too: see Channels and Direct messages.

The New Thread page

A workspace opens on the New Thread page, at the workspace's own address. It greets you by your first name (Good morning, Good afternoon or Good evening) beside the SPARK mark, drawn in the workspace's colour, holds the composer, and lists Recents under it: the threads, canvases, tasks and other items you opened last.

To come back to it, click New Thread at the top of the Threads panel, or press ⌘⇧O on a Mac (Ctrl+Shift+O elsewhere).

Send your first message

1. Type your message in the composer. Shift+Enter starts a new line.
2. Send it with Enter, or with the arrow at the composer's right end (Send message).

A message that names no agent goes to Spark, the built-in agent (see The Spark agent): the composer reads Ask Spark or name an agent… until you name someone. Where Spark is switched off, it reads Call an agent by name to start exploring… and you name the agent yourself.

The thread opens on its own page, in your direct message with the agent that answers it, and is named after its first message. You find it again under that agent in Direct Messages, and on the Threads page.

Name an agent

To ask a specific agent:

• Type @ to open the list, then pick the agent. Agents come first, then a few items of each kind.
• Or start typing an agent's name without the @. The matching agent shows over the composer; press Tab to take it.
• Or open the composer's + menu, then Agents, and pick one.

Once an agent is named, the placeholder changes to How can I help you today?. You can also start from the agent itself: click it under Direct Messages to post in your DM with it, or pick it on the Agents tab of New message to come back here with it named (see Direct messages).

Attach files

Use the + button in the composer:

• Select files picks existing files from the workspace.
• Upload files uploads new ones from your computer.

You can also drag files from your computer and drop them anywhere on the New Thread page, or into an open thread. Each upload shows its progress.

You can attach up to 10 files to a thread, 20 MB in total; the info icon next to Upload files shows the limit. Supported types and their size limits:

The rest of the + menu adds who and what the thread starts with: Agents names one, and Tasks, Canvas, Data and Spaces each open a short search, where a click adds an item or takes it out again. The menu stays open, so you can pick several.

Give it context

A thread keeps the items it is about, its context, and every agent in it is given them on every turn. Start the thread from an item's page, tag items with @, or pick them in the composer's + menu: see Add context to a thread, and Thread context for how it works.

What the agent does

An agent works from its instructions and what you gave it. It calls Tools to take actions and loads Skills for guidance. While it works, its author line reads Thinking… and a live window shows its steps. When it finishes, the steps fold into a badge on the author line, such as 3 steps, before the answer's time.

Click the badge to unfold the steps: each tool call or skill on its own line, with how long it took.

Click a step to open it in a dialog. A tool call's dialog has Input and Output tabs, each scrolling on its own: what the agent called the tool with, and what came back. The copy button at the right of the tabs copies the open tab's text. A skill's dialog shows the skill as the agent read it.

Answers can include charts drawn from your data, in the answer itself. A chart only shows to someone who can read the data behind it.

When an agent searched the web for its answer (see Web search), Sources under the answer lists the pages it used, each a link.

When the agent asks you

Some steps pause the thread and wait for you:

• A guardrail asks you to approve a tool call. Choose Approve, Edit, Reject or Respond for each tool, then Submit.
• The agent asks you a question with a few options. Pick one, or choose Something else and write your own answer.
• The agent wants to read a skill it was not given. Approve or reject it.
• The agent wants to read a whole document and asks how it should read it.

The thread continues as soon as you answer. In a thread several people write in, only the person the agent is answering decides.

Stop an agent

While an agent works, the composer's send button becomes Stop. Click it to halt the run. The answer stays in the thread marked Stopped, with Try again in its toolbar, and your next message gets a fresh answer.

Long answers keep running

The work happens on the server, so you do not have to wait on the page. Move elsewhere, or refresh, and come back: the answer is either still arriving or complete, with its steps kept. In a channel, the thread's line reads Spark is replying (or the name of the agent at work) until the answer is in.

Long threads keep their context

A long thread does not lose its start. The newest turns go to the agent in full. Older turns are summarised once, turn by turn, keeping what was asked, the figures and identifiers, and the tools used, and the summaries go with every turn in their place. When a summary lacks a detail, the agent reads that turn back in full.

Rate an answer

Rest the pointer on an answer to show its toolbar at the end of its author line. Use the thumbs up (Good response) or the thumbs down (Bad response):

• A thumbs up is saved at once.
• A thumbs down first asks what went wrong: Wrong or made-up data, Did the wrong thing, Didn't follow my instructions, Incomplete answer, Too slow, Hard to read or Something else, and an optional comment. Send saves it; Cancel saves nothing.

Once you rate an answer, only the thumb you gave shows; click it again (Remove rating) to take the rating back. Ratings appear in the Feedback section of Monitoring.

The same toolbar copies the answer (Copy) and shows what it took: rest the pointer on the token count to see the model, its provider and, for NGT Auto, the model the turn was routed to, then the tokens used.

Find past threads

Threads are saved as you go. Open Threads at the top of the Threads panel to see every thread in your channels and direct messages, the one written in most recently first, each under where it is. A row shows the thread's first message, who replied, how many times and when. Click it to open the thread beside the list.

To find a thread by what it says, use Search in the sidebar (⌘F on a Mac, Ctrl+F elsewhere) and look under Threads. The New Thread page lists the threads you opened last under Recents too.

Where to go next

• Channels
• Direct messages
• Add context to a thread
• Threads and channels
• The Spark agent
• Create an agent
• Guardrails