UFO³ Galaxy
Coordinates Windows, Linux, and Android task workflows through adaptive DAG-based orchestration.
Per-dimension scores and reasoning
Evidence: README mentions AIP protocol as 'secure coordination layer' but lacks specific security mechanism details; configuration examples require API keys but do not explain least privilege; no user confirmation mechanism or data flow transparency specifics found; dependency list is comprehensive but no vulnerability scanning or security audit evidence; external effects (cross-device operations) mentioned but security controls not detailed; no rollback mechanism found; source attribution is clear (Microsoft official). Deductions: lack of concrete security implementation details, user confirmation, data flow transparency, sensitive data handling are only superficially mentioned without evidence.
Evidence: README and documentation structure are consistent, UFO² and Galaxy positioning is clear; dependency list is complete but no version compatibility notes; failure messages: test files include error handling tests but no user-visible error message examples. Deductions: dependency availability not verified, failure message evidence insufficient.
Evidence: README clearly distinguishes UFO² and Galaxy use cases, provides migration path; capability boundaries explained via comparison table; trigger precision: task decomposition and DAG orchestration described but no specific trigger conditions; environment fit: supports Windows, Linux, Android but no detailed configuration guides. Deductions: trigger precision evidence insufficient, environment fit details limited.
Evidence: README structure is clear, includes quick start, documentation links, FAQ; installation steps simple; naming stable (UFO, UFO², UFO³); examples and FAQ abundant; known limitations mentioned in FAQ (e.g., Galaxy in active development); license is MIT; version history mentioned in updates but no detailed changelog; maintenance responsibility clear (Microsoft official). Deductions: version changelog not detailed.
Evidence: output usability: final reports and status updates provided; marginal value: UFO³ introduces multi-device orchestration, innovative; cost-benefit: mentions '51% fewer LLM calls' but no specific cost analysis. Deductions: cost-benefit evidence insufficient.
Evidence: README cites arXiv papers, provides citation formats; multiple sources (papers, docs, videos) corroborate; facts and inferences distinguished (e.g., 'Formally verified correctness' is a claim but no verification details). Deductions: some claims lack independent verification, but overall traceability is good.
- Static review cannot verify actual runtime behavior; all security claims need further testing.
- Dependency list lacks vulnerability scanning results; recommend checking for known vulnerabilities.
- User confirmation mechanism not explicit; cross-device operations may pose risks.
What does this agent do, and when should you use it?
UFO³ combines the Galaxy multi-device orchestration framework with UFO² Desktop AgentOS for cross-device, cross-application automation. Galaxy uses ConstellationAgent to decompose requests into a TaskStar dependency DAG, while TaskOrchestrator coordinates asynchronous execution and can revise the graph from execution feedback. Device agents register, receive work, and return status through the WebSocket-based AIP protocol; the README explicitly lists Windows, Linux, and Android support. UFO² is the stable Windows automation component and can also act as Galaxy’s Windows device agent, using Windows UIA, Win32, WinCOM, and hybrid GUI/API actions. It is started through local Python commands and requires model API credentials and device-pool configuration rather than functioning as a hosted chat product.
In Galaxy mode, python -m galaxy --interactive starts an interactive workflow in which ConstellationAgent turns a request into a TaskConstellation/DAG, selects devices using platform, resource, task-requirement, and performance-history criteria, and lets TaskOrchestrator schedule dependency-ready work asynchronously or in parallel. Device agents exchange tasks, status, and results through the AIP WebSocket layer; the workflow exposes real-time status, error recovery, and a final report. Windows tasks can be executed by UFO², which uses Windows UIA, Win32, WinCOM, hybrid visual/UIA detection, and a combination of GUI clicks and API calls. Standalone UFO² runs a specified task with python -m ufo --task <task_name>.
- An automation team coordinating a dependency-driven workflow across Windows, Linux, and Android devices.
- A team with existing UFO² Windows automation that wants to use it as a Galaxy execution node while expanding incrementally to multi-device workflows.
- An engineering group that needs to decompose complex requests into parallelizable DAG tasks and revise subsequent work from execution results.
- A developer automating Windows desktop applications with GUI actions, native Windows control interfaces, and API calls.
- A developer building device agents that use MCP for tool augmentation and connect to a shared orchestration layer through AIP.
What are this agent's strengths and limitations?
- Models task dependencies as a Constellation DAG and supports asynchronous, parallel cross-device execution rather than only a single-device sequential ReAct flow.
- UFO² can run independently or serve as Galaxy’s Windows device agent, enabling incremental migration.
- The Windows path explicitly integrates Windows UIA, Win32, and WinCOM, with hybrid GUI and API actions.
- AIP supplies WebSocket-based device-agent coordination, and the README explicitly describes an MCP-based tool-augmentation path.
- Galaxy is labelled Active Development; the README recommends it for experimentation and non-critical workflows, while advanced device types, monitoring, and performance optimization remain in development.
- Multi-device deployment requires device-pool configuration and platform-specific server/client and, in some cases, MCP-server startup, making it more involved than standalone UFO².
- Operation depends on LLM API credentials and network access; the documented examples use OpenAI or Azure OpenAI endpoints.
- The main README does not provide complete startup commands for Windows, Linux, or Android device agents, so deployment requires the platform-specific guides.
How do you install or deploy this agent?
Python 3.10 or 3.11 and usable LLM API credentials are required. Install dependencies with pip install -r requirements.txt. For Galaxy, copy config\galaxy\agent.yaml.template to config\galaxy\agent.yaml and add model settings; the README example uses API_TYPE: "openai", API_BASE: "https://api.openai.com/v1/chat/completions", API_KEY, and API_MODEL: "gpt-4o". Then register devices in config\galaxy\devices.yaml. For UFO² alone, copy config\ufo\agents.yaml.template to config\ufo\agents.yaml and supply equivalent model credentials.
How do you use this agent?
For Galaxy, configure ConstellationAgent and config\galaxy\devices.yaml, start the appropriate device agents, then run python -m galaxy --interactive. The README says Windows needs a server and client, while Linux and Android need a server, MCP servers, and a client, but it does not provide the complete device-agent startup commands. For UFO², configure config\ufo\agents.yaml, then run python -m ufo --task <task_name>.
How does this agent compare with similar options?
Compared with UFO², UFO² targets one Windows device, sequential ReAct execution, and easier setup; UFO³ Galaxy targets cross-device task dependencies, DAG planning, and parallel execution. The README labels UFO² as LTS and Galaxy as active development. For a simple Windows-only automation task, UFO² is the more direct choice; for complex multi-device workflows, Galaxy can orchestrate UFO² alongside other device agents.