Every primitive is the physical embodiment of a claim from the paper, and every specimen imports the shipped component, so the catalogue cannot drift from what the registry ships. Interactive, shadcn-compatible, copy-paste ready.
20 primitives · 17 specimens · MIT
The moments where the human answers. Interrupt boundaries, binary approvals, question sets, queues, and the plan the agent has not run yet.
Human-in-the-loop interrupt cards rendered inline in a chat thread. Three semantic variants, each with idle, expanded, confirmed, and dismissed states, and every resolution can be undone. Click any card to expand it.
The core decision row used across review, download, and notes panels. Three answers, not two: approve, reject, and can't tell, because an unresolved question is not a no. Undo returns to pending.
Multi-question approval card: single choice, multi-select, and a freeform text field. Submits to a resolved line that keeps the answers visible and can be reopened.
Sequential approve-reject flow across mixed agent items. Auto-advances to the next item, can step back one, and resolves to a summary state.
The multi-step plan the human can rename, reorder, add to, or delete from before the agent executes it. Steps marked locked cannot be removed.
What the agent is doing and what it is about to do. Execution states, the reasoning trace, the pending tool call, and the context it was handed.
Six discrete agent execution states. The running state animates. Use it in any card that wraps an in-progress agentic task.
Research Agent
Climate Policy workspace
Research Agent
Climate Policy workspace
Research Agent
Climate Policy workspace
Research Agent
Climate Policy workspace
Research Agent
Climate Policy workspace
Research Agent
Climate Policy workspace
Step-by-step thought, action, result renderer; each step collapses to reveal its detail. A visible implementation of the supporting-facts requirement from §3.3 of the paper.
The tool call the agent wants to make: name, arguments, optional rationale and signals, shown before it runs so the human can approve or reject. Pairs with the gates layer for confidence, cost, and scope checks.
Drafted reply to the client thread; confirming high-stakes outbound before sending.
Pill chips for the context attached to an agent run: notes, files, URLs. Removable, with overflow truncation built in.
What the agent found, and where it came from. Ranked results, source-backed claims, and the exact text a proposed edit would change.
Ranked result cards with metadata, snippet, and a relevance bar. The relevance figure is a signal for the human to weigh, not a verdict the agent has already acted on.
Carbon Pricing Mechanisms and Emissions Outcomes
Stavins, R., Stowe, R., Comstock, M., Nature Climate Change, 2023412 citations
Empirical analysis of 40 carbon pricing schemes across 30 jurisdictions reveals that price levels above $50/tCO₂ are associated with meaningful emissions reductions in the power sector...
Just Transition Frameworks in Coal-Dependent Regions
Newell, P., Mulvaney, D., arXiv, 202487 citations
A systematic review of 28 coal phase-out programmes finds that regions with co-designed transition plans achieved 34% higher re-employment rates within 24 months...
Net Zero Pledges: Verification and Accountability
Höhne, N., Gidden, M., den Elzen, M., Science, 2023301 citations
Of 196 national net zero pledges examined, only 31% include robust interim milestones and independent verification mechanisms aligned with 1.5°C pathways...
EU Emissions Trading System Post-Reform Price Dynamics
Borghesi, S., Flori, A., Energy Economics, 2024156 citations
The Market Stability Reserve mechanism introduced in 2019 has demonstrably reduced permit surplus, contributing to a tripling of EUA prices from 2018 to 2023...
A single source-backed citation: the claim on top, the source attribution below, an expandable supporting quote, and an optional confidence badge. Verify, reject, or can't tell.
Where a claim is grounded, not merely that it is. One row per pointer with the source, the locator in human units, and the excerpt. Sources the agent consulted and drew nothing from are listed too, so silence never reads as safety.
The comment addresses the recipient by name and repeats a threat made in an earlier post.
…you know exactly where I'll be waiting for you, Dana.
I'll be waiting outside after the meeting.
Not assessed: Direct messages between the accounts, Reports from other users
Before and after for a proposed text or code edit, with per-hunk strips. Drop it into any agent loop where the human should see exactly what will change before it lands.
It seems like the central question of this paper might be how AI systems should be evaluated when they are deployed in collaborative settings.This paper argues that AI systems must be evaluated against collaborative performance, not autonomous task completion.Whole task surfaces built from the primitives above. The shape a real agent panel takes once the parts are assembled.
A compound widget for a draft in progress: title, target section, word range, evidence notes, and the same six status states the subagent card uses.
Evidence notes
Three operating modes for a long-running research task: create a new session, follow up on an existing one, or read a single URL.
How much of this did a person do. One five-point scale from Human to AI in four densities, slider to inline pill, plus the five accents and four approval badges the kit draws from.
One question, how much of this did a person do, answered on a five-point scale from Human to AI. Every density shows the same thing: a track filled to the current level, revealing the spectrum from emerald to rose as it goes, and the level in plain words. Colour is the cue; the words carry the meaning. Labels sit on foreground and muted-foreground, never on colour.
Written together, turn by turn.
Reach for this when the person sets the value. The readout says the level and what it means, and the track fills as far as the thumb, revealing the spectrum. Drag it, tap either end, or focus it and use the arrow keys, Home and End.
Reach for this to show provenance in a list row or a header without inviting interaction. Read-only by design: one image element with no focusable children, so fifty rows do not become fifty tab stops.
Reach for this in a table cell or a queue row where even the meter is too much furniture. Static by default; given an action handler it grows ‹ › steppers with real 24px targets that go inert at the ends of the scale without dropping keyboard focus.
CollaborativeWritten together, turn by turn.
The explicit form, for a settings panel or a form where every option should be visible and directly tappable. Five segments in one pill, the chosen one raised, each carrying its colour so the order reads before the words do. It needs about 400px; below that, use the slider.
Each primitive installs on its own through the shadcn CLI. No fork, no wrapper SDK. The registry page carries the exact command and dependency chain for every item.