Acerca de
October 2026 EMNLP system demonstration separating reusable graph, table, text, and display views from task-specific interaction meanings. ABRA publishes these bindings as shared canvas state so users and other agents can inspect and invoke task-specific actions.
Resumen
Why ABRA Adds New Signal:
1. Reusable interaction vocabulary: Adapts what an action means through explicit task bindings
2. Shared semantics: Other agents can inspect and invoke published actions; local view effects and owner-agent consequences follow separate paths
3. Reference implementation: React views, a Node.js canvas hub, and Python adapters support shared human-agent workspaces
4. Evaluation boundary: Four models are tested on isolated one-turn operations; exact view selection reaches only 33–53%, with user understanding and end-to-end task benefits untested
[Official prototype, setup instructions, and demo](https://github.com/megagonlabs/abra).
1. Reusable interaction vocabulary: Adapts what an action means through explicit task bindings
2. Shared semantics: Other agents can inspect and invoke published actions; local view effects and owner-agent consequences follow separate paths
3. Reference implementation: React views, a Node.js canvas hub, and Python adapters support shared human-agent workspaces
4. Evaluation boundary: Four models are tested on isolated one-turn operations; exact view selection reaches only 33–53%, with user understanding and end-to-end task benefits untested
[Official prototype, setup instructions, and demo](https://github.com/megagonlabs/abra).
Etiquetas
generative-uishared-canvasesinteraction-semanticsdirect-manipulationmulti-agent