PhyAgentOS/PhyAgentOS-core

PhyAgentOS is a Recursive Self-Improving (RSI) physical agent operating system that enables agents to recursively self-improve through agentic workflows.

What it solves

PhyAgentOS provides a structured infrastructure for managing embodied AI agents that interact with physical robots or simulators. It solves the problem of unreliable robot execution by separating high-level planning from low-level execution and introducing a rigorous verification system that uses visual and state evidence to determine if a goal was actually achieved, rather than blindly trusting execution logs.

How it works

The system operates through a decoupled architecture:

  • Agent Loop & Planner: The high-level AI agent plans tool calls to achieve a goal.
  • Forge Gateway: A versioned API boundary that executes robot actions. The agent never interacts directly with hardware SDKs or simulators.
  • Evidence Collection: The system captures "before" and "after" images and robot states around actions to provide ground-truth evidence.
  • Task Verifier: A separate component that evaluates the goal, success criteria, and captured evidence to issue a verdict (success, failure, or replan required).
  • Experience & Evolution: Successful and failed tasks are recorded in SQLite databases. The system can asynchronously reflect on these tasks to generate "Lessons" or promote new Skill revisions to improve future performance.

Who it’s for

It is designed for developers and researchers working on embodied intelligence, robotics, and AI agents that require verifiable, crash-safe execution in physical or simulated environments.

Highlights

  • Evidence-Based Verification: Uses visual and state evidence to verify goals independently of the action taken.
  • Recursive Self-Improvement: Automatically generates scoped lessons from repeated failure patterns and promotes validated skill revisions based on successful episodes.
  • Crash-Safe Orchestration: Uses SQLite WAL for persisting task logs, plan revisions, and execution records.
  • Decoupled Execution: A strict Forge Gateway contract ensures the agent remains agnostic to the underlying robot or simulator implementation.
  • Managed Skill Runtime: Supports versioned, SHA-256 verified skill bundles and independent installation/management of robot capabilities.

Related

  • Project
    fy-agent/fyagent

    FyAgent is a cross‑platform desktop app that consolidates the configuration of AI‑assistant tools (model/provider selection, skill/plugin installation, prompt and memory editing, and service credentials) into a single local UI, helping users manage multiple AI agents consistently.

  • Project
    ZJU-REAL/HugAgentOS

    An enterprise-grade, self-hosted AgentOS that uses domain ontologies to govern AI agent reasoning and execution for trustworthy, tool-augmented task completion.

  • Project
    HorizonRobotics/HoloAgent

    A unified embodied-agent framework for general-purpose robots that integrates 3D spatial memory and closed-loop execution for long-horizon real-world tasks.

  • Project
    compozy/compozy

    An operating system for AI agents that provides a daemon-based runtime to manage persistent sessions, automation, and memory for ACP-compatible agent CLIs.

  • Project
    zai-org/Synapse

    A self-hosted AI workspace that transforms chatbots into digital teams by providing a governed runtime for humans, AI actors, and remote agents to collaborate in shared conversations.