takahirom/arbigent

AI Agent for testing Android, iOS, and Web apps. Get Started in 5 Minutes. Arbigent's intuitive UI and powerful code interface make it accessible to everyone, while its scenario breakdown feature ensures scalability for even the most complex tasks.

What is Arbigent?

Arbigent (short for Arbiter‑Agent) is an open‑source framework for testing AI‑driven agents that interact with user interfaces. It lets you write high‑level, natural‑language goals (e.g., “log in and search for a video”) and then runs those goals on real or emulated devices (iOS, Android, web, TV). The framework breaks a complex goal into smaller, dependent scenarios, orchestrates the AI’s actions, and checks the results with image‑based assertions.

Why it matters

Traditional UI tests are brittle – a tiny layout change can break a script. Because an LLM can understand the UI, Arbigent’s tests are more resilient: the AI can adapt to minor UI changes, reducing the maintenance burden. It also brings AI‑agent testing into the normal QA workflow with a visual UI for non‑programmers and a code‑first interface for automation pipelines.

Core features (as described in the README)

Category Highlights
Task decomposition Scenarios can depend on each other (e.g., login → search). The framework orchestrates the flow.
Hybrid workflow UI for creating scenarios; YAML files for programmatic execution (arbigent run …).
Cross‑platform Supports iOS, Android, web and TV (including D‑pad navigation).
AI optimisation UI‑tree filtering, annotated screenshots, and optional AI hints embedded in accessibility labels to improve model understanding.
Cost‑effective models Defaults to gpt‑4.1; also works with cheaper models like gpt‑4o‑mini.
Reliability aids AI‑powered image assertions (via Roborazzi), stuck‑screen detection, retry and replay‑with‑fallback caching.
Extensibility Custom hooks, Model Context Protocol (MCP) servers for external tools, and reusable scenario library (similar to GitHub Actions).
Maestro YAML integration Run existing Maestro test flows as initialization steps, letting you set up deterministic app state before the AI takes over.
Open source Free to use, modify and distribute; encourages community contributions.

Who would use it?

  • QA engineers who want a point‑and‑click UI to define tests without writing code.
  • Developers looking to embed AI‑agent tests into CI pipelines via the CLI and YAML configuration.
  • Product teams that need to test complex, dynamic flows (e.g., onboarding, video playback) where traditional scripts flake.
  • Companies that already use Maestro or other UI automation tools and want to add an AI layer without discarding existing assets.

How to get started (per the README)

  1. Download the UI binary from the GitHub releases page.
  2. Connect a device (real or emulator) and enter your AI provider’s API key in the app.
  3. Create a scenario in the UI by typing a natural‑language goal.
  4. (Optional) Import existing Maestro YAML or Android Studio Journeys files to set up app state.
  5. Run the scenario either from the UI or via the CLI, e.g.:
    arbigent run --project-file=project.yaml --scenario-ids=login-search
    

Limitations noted by the author

  • Speed – limited by LLM latency and real‑time UI interaction (≈ 1‑2 seconds per step). Parallel execution (--shard) and result caching can mitigate this.
  • Cost – each AI step incurs API charges (e.g., $0.005 per step with GPT‑4o).
  • Reliability – still depends on device stability; emulator flakiness can cause failures.
  • Prompt engineering – may be required for complex apps, though the goal is for models to improve over time.

All information above is taken directly from the repository’s README; no additional features have been inferred.

Related

  • Dispatch
  • Project
  • Project
  • Project
  • Project