Thread context
What a thread is about: the items it holds, how they get there, what agents do with them, and how a space narrows a thread.
A thread's context is the set of items it is about: the data objects, files, canvases, tasks, workflows and their runs, spaces and agents the conversation works with. It is the thread's working set. It shows beside the thread in the context panel, and every agent that answers in the thread is given it on every turn.
Why it matters
An agent does its best work on the right few things, and the context says which. When you write "the file" or "that canvas", the agent takes it to mean an item in the context, and it looks there before it searches the workspace. A thread you start from a canvas is about that canvas from its first message.
Pointers, not copies
An item in the context points at the item; it is not a copy. On each turn the agent is given each item's kind, name and a line about it, never its content, and it opens what it needs with its tools. So it always reads the item as it is now: a record changed this morning reads as changed, and a renamed canvas shows its new name. An item deleted or archived since it was added is marked unavailable, and the agent says it could not read it rather than guess.
How items get in
You add them:
- Start from an item. The new-thread button on a task, canvas, data object, workflow or space starts a thread with that item in its context.
- Tag it. Type @ in a message and pick the item.
- Pick it. Open Tasks, Canvas, Data or Spaces in the composer's + menu, or use the + in the context panel.
- Send a file. Every file attached to a message joins the context.
Agents add them too:
- What an agent creates in the thread joins on its own: a canvas it builds, a task it opens, a workflow it writes, a space it puts items in.
- What you are about to work with. When a search turns up the item you clearly want next, the agent may attach it so it stays for later turns. It attaches the few things that matter, not everything it looked at: what it only listed or read stays out.
An item leaves when you take it out, with Remove in its menu in the context panel. An agent takes an item out only when you say you are done with it. A thread holds up to 50 items.
A space narrows the thread
A space in the context brings its items along and sets the thread's scope. The agent works with the space's items: asked "what data objects do we have?", it answers from the space rather than the whole workspace, and says so. To widen the thread again, ask the agent to look beyond the space, or take the space out of the context. See Spaces.
What an agent knows in a thread
The context is one of five things an agent answers from:
| Source | What it gives the agent | How long it lasts |
|---|---|---|
| Its instructions | Who it is, what it does and what it must not do | Every thread it answers in |
| Its skills | How to do a kind of task, read when the task calls for it (see NGT Skills) | Every thread it answers in |
| Its memory | What it saved to recall in later threads (see Memory) | Across threads |
| The thread's context | Which items this conversation is about | This thread |
| The conversation | What was said and done so far. Older turns are summarised, and read back in full when a detail is missing | This thread |