ResearchResearch paper

Interface Quilt turns computer-use agents and generative UI into persistent patches for existing applications

Interface Quilt from Autodesk Research combines computer-use agents with generative UI. Agents first explore controls and observable state through rendered pixels, then a coding agent generates an interface patch over the existing application. The paper demonstrates reordered controls, voice and marking-menu input, in-situ parameters, and delayed or confirmed sends across PowerPoint, Affinity, Fusion, Slack, and Overleaf without source code, application APIs, or accessibility data.

Why it matters

The work moves generative UI from a one-off answer toward a reusable application-modification layer. The generated interface continuously reads host state and translates interaction back into host actions, making pixels, synchronization, OS input, and generated-code permissions part of the product boundary.

A closer look

The system separates build time from runtime. OmniParser v2, Apple Vision OCR, and Holo2-8B identify controls and coordinates before Claude Code generates TypeScript patches. On macOS, a transparent Electron surface hosts patches, a Node controller dispatches events, and a Swift daemon performs OS input and exposes local perception services. Patches can keep sliders, situated menus, or usage traces synchronized with the application and map patch actions back to original controls; some examples also call external speech, transcription, or analysis models.

Who this is for

Product, design, and engineering teams studying malleable software, computer-use agents, desktop automation, assistive interaction, and generative-UI runtimes.

Original sources

Revisions & corrections (1)

— Backfill a newly discovered primary research source directly about generative UI and computer-use agents, preserving the real submission date, system architecture, demonstrations, and explicit privacy, security, synchronization, and evaluation limits.

Related events

Research

ControlAnything lets a coding agent generate spatial controls and physical-device connectors on demand

ControlAnything lets a Meta Quest 3 wearer point at an object and describe what their hands should control. A Muse Spark coding agent on a laptop then writes, tests, and deploys a MIDI, Bluetooth, or key-press connector and returns AR knobs, keys, touch surfaces, or a per-frame hand program. The paper was submitted November 5, and the open repository includes real-device runs, an emulator, evidence records, and replayable paths.

Research

EnvOS audits computer-use agents when the world has changed

Submitted October 29, EnvOS archives interfaces, tools, state mutations, observation delivery and scoring evidence to test whether agents recheck facts after appointments, policies or other world state has changed instead of relying on stale memory or prompts.

Research

A second, independent VeriCUA turns GUI-task acceptance into active evidence gathering with abstention

A Google and University of Hong Kong team submitted a separate project also named VeriCUA on October 29. It inspects an environment or recording before constructing a task rubric, then gathers missing evidence through GUI, accessibility-tree, file and database reads, marking requirements met, unmet or unknown. In online experiments it reduces false positives against baselines on AndroidWorld, OSWorld and ScienceBoard.