You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Status: Draft — open for input · Category: Big bet
Summary
A single queue of everything agents are waiting on you for: paused runs needing input, corrections awaiting a decision, PRs to review. One place to triage instead of hunting across screens.
Why
As automation grows, the bottleneck shifts from "can an agent do it" to "did a human notice it needs them." Scattered signals get missed; a unified inbox makes the human-in-the-loop part reliable.
Sketch
Aggregate actionable items (HITL pauses, pending corrections, open improvement PRs) into one workspace inbox.
Each item links to its full context and the action that clears it.
Filter by agent, type, and assignee. (Provisional — the shape changes based on the discussion.)
Open questions
What counts as a "task" — just blocking items, or also FYIs?
Per-user vs per-workspace queues?
Notifications: in-app only, or email/Slack too?
How to help
React 👍 if you want this, and comment with your use case, constraints, or pushback. This is a draft Tembo Agent Studio Improvement Proposal (TASIP); your input shapes it before it becomes scheduled work.
✅ Shipped — the Tasks Inbox is built and live on main
This went from draft to shipped over the v0.4 cycle, driven by a real first consumer (a LinkedIn inbox-triage agent). It's deployed to the internal instance.
What shipped
The substrate — one workspace queue (inbox_item) that humans and agents read and write. Source-agnostic: each item carries source, external_ref (idempotent upsert), item_type, context, a proposed action, and a status lifecycle (open → claimed → awaiting_human → done/dismissed). (4e1797a)
Agent-actionable API — the inbox is itself a tool surface. produce_inbox_item, list_inbox_items (search / filter / sort), and claim / propose / complete are exposed over both Native-MCP and /api/v1/inbox REST, so agents produce and act on items as peers to humans. (4e1797a, 59ad790)
Action menus + on-click execution — an item can declare a set of typed options (reply / one-click), one marked recommended; the inbox renders them as buttons and runs a synchronous executor on click (e.g. send / archive on the source system). (59ad790)
Batched learning loop — each item's (proposed vs. final) diff is a learning signal. Agents in "learning mode" aggregate signals and self-update on a schedule — one improvement PR per cycle, not one per signal. (4e1797a)
UI — Inbox at the top of the sidebar with a live active-count badge (now revalidated the moment you act on an item), friendly context rendering, search / filter / sortable headers via a shared DataTable, success toasts, and snooze (move an item out until later; it reopens automatically — or sooner if a newer reply lands). (b080e6b, d3d6cc1, 8d71a71, 93d22da)
First consumer — LinkedIn triage — keeps the queue capped at N open threads (top-up, not fetch-N), skips archived / handled / snoozed threads, paginates back for fresh ones, and drafts replies with one-click Send / Send + Archive / Archive. (50e97ae, 857d147, 9d09a9a, 1b78ab9)
Original open questions — resolved
What counts as a task? Anything an agent surfaces for a human, blocking or not. The producer decides; status + snooze keep FYIs from clogging the queue.
Per-user vs per-workspace? Per-workspace queue (single-tenant deployment), with assignee fields on each item for future per-user routing.
Notifications? In-app (sidebar badge) for now; email / Slack deferred.
Closing as shipped. Follow-on polish and the HITL-pause source are tracked in the parity epic (#169).
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Status: Draft — open for input · Category: Big bet
Summary
A single queue of everything agents are waiting on you for: paused runs needing input, corrections awaiting a decision, PRs to review. One place to triage instead of hunting across screens.
Why
As automation grows, the bottleneck shifts from "can an agent do it" to "did a human notice it needs them." Scattered signals get missed; a unified inbox makes the human-in-the-loop part reliable.
Sketch
(Provisional — the shape changes based on the discussion.)
Open questions
How to help
React 👍 if you want this, and comment with your use case, constraints, or pushback. This is a draft Tembo Agent Studio Improvement Proposal (TASIP); your input shapes it before it becomes scheduled work.
All reactions