Affora: A Design System for Agent-Friendly Interfaces
cs.HC · cs.AI · cs.SE
Quick verdict
A design system exposing a machine-readable semantic substrate for shared human-agent interfaces; broad but author-built evidence, statistically untested.
Summary
The paper presents Affora, a design system for web interfaces that serve both human users and computer-use agents. Its core construct is a 'semantic substrate' - controls, names, relationships, state, choices, and outcomes recoverable from the rendered interface - which must stay explicit while visual expression varies. Three studies ground the rules: (1) across 60 components, native HTML reached 91% agent success versus 68.3-86.7% for eight component libraries, and a repair sequence lifted success from 43% to 90%; (2) with the substrate fixed, sixteen themes and six layout archetypes preserved ~99.5% DOM-channel completion, and a signed visual-cue sweep showed no monotonic benefit from reinforcing human affordances; (3) seven canonical interaction-design principles transfer unevenly - confirmation steps and form step-splitting added interaction cost without completion gains. Evaluations on WebArena's Magento, WebShop, and four independently authored applications show Affora reaching 100% completion versus a 77% baseline, with gains concentrated where interfaces hide targets or state; ARIA-only augmentation did not consistently help, and an agent.md instruction file matched Affora on the primary set. A release-workflow case showed 21-40% fewer actions and 24-41% fewer tokens, from only 1-2 successful pairs per condition. Key limitations: no statistical inference, author-built harness and tasks, web-only scope, no human-subjects evaluation, and an unvalidated executable check suite.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Agent failures on ordinary components are substantially explained by the interface's semantic substrate rather than model capability: repairing semantics and exposing choices/state lifted success from 43% to 90% (Study 1).
- When the semantic substrate is fixed, substantial visual variation (16 themes, 6 layouts) preserves agent completion (~99.5% DOM channel), so visual freedom and agent reliability are compatible (Study 2).
- Canonical interaction-design principles transfer unevenly to computer-use agents: confirmation steps and multi-step forms add interaction cost without demonstrated completion benefit (Study 3).
- Affora-style rewrites improve completion mainly where a targeted deficit exists (WebShop hidden-radio: 8% to 68% after extending one rule), with little or no effect on already-sound interfaces.
- In an author-built release workflow, Affora reduced actions by 21-40% and total tokens by 24-41% across four model-channel conditions, based on only 1-2 successful pairs per condition.
Strengths
Causal deficit manipulation, not just before/after correlation
The Magento three-model deletion-restoration experiment (Table 3) introduces a targeted deficit (completion 15% to 7%) and then repairs it (7% back to 16%), isolating the specific rule's effect; 287 unaffected task pairs show no systematic movement, a genuinely experimental design rare in this space.
Unusually honest evidence bookkeeping
The appendix principle index labels every rule as measured, observed, null, or proposed; the text repeatedly bounds claims (workflow case is 'preliminary evidence', WebMCP is 'not a matched comparison'); and negative results are reported plainly (MUI and Atomic CRM near-null at 79%->81% and 72%->71%, shadcn-admin predictably-labelled group drops 79%->74%).
Study 1 cleanly localizes failure to the interface layer
A useful, well-bounded null result on visual freedom
Comparison against alternative agent-facing mechanisms
Weaknesses
No statistical inference anywhere
Self-built evaluation stack with no external check
Instruction-file parity undercuts the core thesis
Restoration effect operates at very low absolute success rates
Headline artifact (executable checks) is unvalidated
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Affora?
Affora is a design system proposed in arXiv paper 2609.19125 for building web interfaces that serve both human users and computer-use agents. It keeps a machine-readable 'semantic substrate' - controls, names, state, choices, and outcomes - explicit and operable in the rendered interface, while letting visual styling vary freely. It ships reference UI components, design rules across component/layout/flow/site scales, and executable conformance checks.
How much does Affora improve agent task success?
In the paper's own evaluations, Affora reached 100% completion versus a 77% baseline on its primary comparison set. Gains concentrated on interfaces that hide targets or state behind non-predictive labels (WebShop hidden-radio tasks rose from 8% to 68% after one rule was extended). On already-sound interfaces (MUI dashboard template, Atomic CRM) effects were near zero, and a plain agent.md instruction file matched Affora on the primary set.
Does designing for agents hurt human users?
The paper does not test this. It argues agent-friendly changes mostly affect the machine-readable layer (overlapping with accessibility practice) while leaving visual design unchanged, but human usability, trust, and supervisability are explicitly left to future work - no human-subjects study was conducted.
Does Affora work outside web browsers (desktop apps)?
Not demonstrated. All evaluations are web-based, using DOM and accessibility-tree observation channels. The authors argue the principles could map to desktop accessibility APIs (Windows UI Automation, macOS accessibility, Linux AT-SPI, Electron), but this transfer is hypothetical and untested in the paper.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-8 could not be analysed from the extracted text; figure-dependent detail (e.g., Study 2 per-condition results, Fig. 5 principle panels, Fig. 8 trajectory) is taken from captions and body text only.
- Only the design-system components are asserted to be 'distributed as source'; the evaluation harness, task specifications, and analysis code have no stated repository or archival link, so 'codeAvailable' reflects the component-library claim only.
- Model names ('ChatGPT 5.6 Luna/Terra') and tool versions are reported as given in the text and could not be independently verified.
- Some results are described selectively in-text (e.g., 'the available Luna marked-pixel run' for the visual-cue sweep), so full per-condition coverage of Studies 2 and 3 cannot be confirmed from the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.