AI design agents need dense, quantifiable feedback to iterate beyond static style transfers. We’re investigating tools that generate structured signals, enabling critique loops that teach agents about functional design quality. Maze is a prime candidate: it’s a user research platform that produces per-screen usability metrics, which can serve as reward signals for an agent’s hill-climbing process.
This review is based on official documentation, pricing pages, and community reports — we did not run the tool hands-on.
Context
Maze is a user research platform for conducting unmoderated, task-based testing of Figma prototypes and live websites. It produces both quantitative metrics and qualitative data from real user interactions (Source: Maze Reporting Guide). For an AI design agent, Maze acts as an oracle, converting subjective usability into a structured data log. It moves feedback from a vague “this feels cluttered” to a precise “screen X had a 45 usability score and 40% misclicks.”
Agent-Perceivable Signals
An AI agent, connected via Maze’s MCP server (available on Enterprise plans) or CSV exports, can ingest the following reward signals:
- Per-Screen Usability Score (0-100): The core composite metric, computed per study, task, and screen from success rate, duration, and misclicks (Source: Maze Reporting Guide). Thresholds: High (80-100), Medium (50-80), Low (0-50).
- Component Metrics: Success rate, average duration, and misclick rate for each screen in the expected user path (Source: Maze Reporting Guide).
- Click Heatmaps & First-Click Data: Provides click density, per-area metrics, and the critical first click, attributing interaction patterns to specific UI coordinates (Source: Maze Click Heatmaps Guide).
- Path Data & CSV Exports: Detailed logs of the path taken, clicks, time per screen, and misclicks, exportable as CSV for analysis (Source: Maze Reporting Guide).
The platform integrates directly with Figma for prototype import and offers a Model Context Protocol (MCP) server that allows AI assistants like Claude and ChatGPT to query study data programmatically. This turns the research platform into a queryable API for an agent’s perception loop.
Analysis
Maze’s power lies in its localization of feedback. Instead of a single score for an entire user flow, it breaks usability down by screen. This is the critical unit of iteration for a design agent. If a task fails, the agent can pinpoint the exact frame where success rate plummeted or duration spiked.
The usability score becomes a dense, multi-objective reward signal. An agent could be tasked to maximize this score for a given flow. The three-component calculation (success, duration, misclicks) is a built-in guard against overfitting. An agent that only optimizes for success rate might render every button a giant, obvious target, degrading aesthetics and efficiency (as measured by duration). A composite score forces a more balanced design optimization—a step toward genuine “design sense.”
The available data enables a classic reinforcement learning pattern: Observation (fetch screen usability scores via MCP or CSV), Action (modify the Figma layout for a low-scoring screen), Experiment (re-run the Maze study), Reward (measure the delta in the usability score). The per-screen attribution makes the reward assignment unambiguous.
Caveat: Maze measures task completion as a proxy for design quality. It’s excellent for usability and function but blind to brand alignment, emotional tone, or visual harmony. An agent optimizing solely on Maze scores could produce efficient but soulless interfaces. The reward function must be multi-dimensional.
What to Learn
For an AI design agent’s feedback loop, Maze provides the structured, quantitative “ground truth” about user behavior that is essential for measurable improvement.
Computable Rule: Screen Usability Threshold Rule — If a screen’s usability score is below 50, flag it as high-friction and prioritize it for redesign. The Empire Design Scan could parse Maze CSV exports or use the MCP server to fetch screen-level usability scores and trigger alerts below this threshold.
The key learning is that good feedback signals must be attributable. Whole-flow scores are too coarse for iterative optimization. By providing metrics per screen in the expected path, Maze enables agents to perform fine-grained, iterative critiques: “This login screen has a 35 usability score due to 60% misclicks—rearrange the form fields and increase the CTA affordance, then re-test.” This turns vague improvement goals into a measurable hill-climbing exercise.
The final design must still incorporate aesthetic and brand goals, which Maze cannot measure. The optimal setup is a multi-tool critique loop: Maze for usability metrics, plus aesthetic judgment models for visual critique, all feeding into the agent’s iterative process.