Blank chat box no be good place to run consequential work. E collapse five very different questions: wetin you want, wetin you dey look at, wetin e base on, wetin you get permission to do, and wetin don already happen, into one stream wey no separate. For casual tasks, dat one fit be okay. For governed operations, where actions touch systems of record and decisions must be defensible, dat collapse na exactly wetin you no fit afford.
Threada workspace dey deliberately split into five surfaces. Each one answer one of those questions, and keeping dem separate na wetin make di work reviewable.
1. Di intent bar: wetin you want?
Work for Threada start from persistent intent bar, no be deep navigation. You state di outcome in natural language, optionally with structured commands, and di runtime turn am into structured, executable artifact: WorkItem with extracted entities, confidence score, and risk flags.
This na intent-first interaction. Instead of forcing operator to know which form, which queue, and which workflow apply before work fit start, di system capture di goal and assemble di path. When information dey missing, e ask for exactly wetin e need instead of showing long static wizard upfront.
2. Di adaptive canvas: wetin you dey work on?
Di canvas na where WorkItem dey live and take shape. E adaptive: di UI fit assemble temporary forms, comparisons, and decision panels to collect missing context and complete di task, instead of rendering one fixed layout for every kind of work.
Generated output default to editable draft, no be committed change. Di operator review, edit, and decide. Control affordances dey explicit: lock and no-change zones, side-by-side compare, fast undo, and version rollback. So di canvas na place to reason, no be place where model first guess become truth.
3. Di evidence drawer: wetin e base on?
Every consequential output suppose fit show im work. Di evidence drawer hold citations, retrieval traces, and source attribution wey ground di WorkItem. When di system no fit ground answer, e talk am clearly with fallback reason instead of inventing confidence.
This surface make “trust di AI” become claim wey person fit inspect, no be blind faith. Operator no need believe draft; dem fit open drawer and check wetin e stand on, how fresh di sources be, and where each claim come from.
4. Di action controls: wetin you fit do?
Reading and drafting safe. Acting on di world no safe by default, so controls surface dey governed. Na there proposals become approvals and approvals become executed actions against external systems: refund, ticket, record update, access grant.
Governance here show as policy: permissions, thresholds, approval gates, and redlines; no be scattered settings toggles. High-risk actions move through explicit proposed, approved, executing progression, and only auto-execute where policy allow am. Service-level kill switch fit halt execution before any connector dey called while preserving state for review. Controls surface na where di system caution become concrete.
5. Di run log: wetin don happen?
Run log na timeline of di WorkItem: every transition, every approval, every action, every AI participant event, in order. Na di surface where receipts gather become history.
Important thing be say AI actions appear as distinct actor events, dem no fold into human activity. When you read run log, you fit know who propose, who approve, and wetin execute, human or agent, without guessing. Run log na wetin auditor read at quarter end and wetin operator read to understand di case in front of dem today.
Why di split na di point
E for simple to build one surface and let everything blur together. Di reason to avoid am na because consequential work demand say these questions stay separate.
If intent, evidence, and action share one surface, e become easy to act on something wey never ground, or approve something wey basis you never see. By giving each one im own surface, Threada make di careful path become di natural path: state intent, shape draft on canvas, check evidence, then act through governed controls, with run log recording everything.
Di five surfaces stay constant across packs and roles; wetin fill dem adapt. Dat stability deliberate. Operator wey learn di shape of one workspace don learn di shape of all of dem, whether dem dey run IT access provisioning, vendor security review, or procurement approval. Di work dey change. Di way you reason about am no dey change.