A blank chat box is a poor place to run consequential work. It collapses five very different questions — what do you want, what are you looking at, what is it based on, what are you allowed to do, and what has already happened — into one undifferentiated stream. For casual tasks that is fine. For governed operations, where actions touch systems of record and decisions must be defensible, that collapse is exactly what you cannot afford.
Threada’s workspace is deliberately decomposed into five surfaces. Each answers one of those questions, and keeping them distinct is what makes the work reviewable.
1. The intent bar — what do you want?
Work in Threada starts from a persistent intent bar rather than deep navigation. You state the outcome in natural language, optionally with structured commands, and the runtime turns it into a structured, executable artifact: a WorkItem with extracted entities, a confidence score, and risk flags.
This is intent-first interaction. Instead of forcing the operator to know which form, which queue, and which workflow applies before they can begin, the system captures the goal and assembles the path. When information is missing, it prompts for exactly what it needs rather than presenting a long static wizard up front.
2. The adaptive canvas — what are you working on?
The canvas is where the WorkItem lives and gets shaped. It is adaptive: the UI can assemble temporary forms, comparisons, and decision panels to collect missing context and complete the task, rather than rendering one fixed layout for every kind of work.
Generated output defaults to an editable draft, not a committed change. The operator reviews, edits, and decides. Control affordances are explicit — lock and no-change zones, side-by-side compare, fast undo and version rollback — so the canvas is a place to deliberate, not a place where the model’s first guess becomes truth.
3. The evidence drawer — what is it based on?
Every consequential output should be able to show its work. The evidence drawer holds the citations, retrieval traces, and source attribution that ground the WorkItem. When the system cannot ground an answer, it says so explicitly with a fallback reason instead of inventing confidence.
This is the surface that makes “trust the AI” an inspectable claim rather than a leap of faith. An operator does not have to believe a draft; they can open the drawer and check what it stood on, how fresh the sources were, and where each claim came from.
4. The action controls — what can you do?
Reading and drafting are safe. Acting on the world is not — so the controls surface is governed. It is where proposals become approvals and approvals become executed actions against external systems: a refund, a ticket, a record update, an access grant.
Governance here is expressed as policy — permissions, thresholds, approval gates, and redlines — not as scattered settings toggles. High-risk actions move through an explicit proposed, approved, executing progression, and only auto-execute where a policy allows it. A service-level kill switch can halt execution before any connector is called while preserving state for review. The controls surface is where the system’s caution is made concrete.
5. The run log — what has happened?
The run log is the timeline of the WorkItem: each transition, each approval, each action, each AI participant event, in order. It is the surface where receipts accumulate into history.
Crucially, AI actions appear as distinct actor events, not folded into human activity. When you read the run log you can tell who proposed, who approved, and what executed — human or agent — without guessing. The run log is what an auditor reads at the end of the quarter and what an operator reads to understand the case in front of them today.
Why the split is the point
It would be simpler to build one surface and let everything blur together. The reason not to is that consequential work demands that you keep these questions separate.
If intent, evidence, and action share a surface, it becomes easy to act on something you never grounded, or to approve something whose basis you never saw. By giving each its own surface, Threada makes the careful path the natural one: state the intent, shape the draft on the canvas, check the evidence, then act through governed controls — with the run log recording all of it.
The five surfaces stay constant across packs and roles; what fills them adapts. That stability is deliberate. An operator who learns the shape of one workspace has learned the shape of all of them, whether they are running IT access provisioning, a vendor security review, or a procurement approval. The work changes. The way you reason about it does not.