Harness, a developer tool released on GitHub on May 3, 2026, uses AI agents to perform authentic user simulation for testing applications. Rather than running predetermined scripted UI tests, the tool has an LLM agent attempt goals described in plain language while flagging UX friction points that actual users would encounter. The project has gained 155 stars on GitHub.
AI Agents Test Applications Like Real Users
Developers can describe testing goals in plain language—such as "sign up and create my first list"—and Harness's AI agent attempts to complete the task. The tool flags UX friction points including ambiguous labels, dead ends, or unresponsive controls that would frustrate real users. This approach differs fundamentally from traditional scripted UI tests that follow predetermined paths.
The tool supports three platforms:
- iOS Simulator: Uses WebDriverAgent for input and control
- Native macOS apps: Uses CGEvent for system-level interaction
- Web apps: Uses embedded WKWebView for rendering and control
Each platform leverages platform-specific input methods and capture techniques to simulate authentic user behavior.
Set-of-Mark Targeting and Credential Management
For web testing, Harness uses a "Set-of-Mark" approach that displays numbered badges over focusable elements in screenshots. This allows the agent to click by ID rather than guessing pixel coordinates. Importantly, the disk PNG saves the clean rendered page—agent scaffolding never appears in exported artifacts.
The tool includes secure credential management where applications can pre-stage username/password pairs. Password bytes never enter the model's context or logs; only metadata indicating a credential was used appears in the system.
LLM Support and Requirements
Harness is compatible with multiple LLM providers:
- Claude (Opus, Sonnet, Haiku)
- OpenAI (GPT-5 Mini, GPT-4.1 Nano)
- Google Gemini models
Each model can be configured with custom token budgets. The tool requires macOS 14 or later and Swift 6, and is released under the MIT license.
The project represents an innovative approach to user experience testing, transforming how developers identify friction before users encounter it in production environments.
Key Takeaways
- Harness uses AI agents to perform authentic user simulation testing for iOS, macOS, and web applications with 155 GitHub stars as of May 3, 2026
- Developers describe testing goals in plain language, and the AI agent attempts tasks while flagging UX friction points like ambiguous labels or dead ends
- The tool uses Set-of-Mark targeting for web testing, displaying numbered badges over clickable elements while keeping exported screenshots clean
- Harness supports Claude, OpenAI, and Google Gemini models with configurable token budgets and secure credential management
- The project requires macOS 14+ and Swift 6, and is released under the MIT license on GitHub