From Virtual Hands to AI for Survivalists: Curious Open Agent OSes [and One Hardware Project]

Among them, for example, is OpenFang, where action execution is handled by "hand" modules, a project that enables managing Linux schedulers, and even a modular field computer with an AI assistant that can be useful in the post-apocalypse wilderness with no connection to the outside world.

Master of All "Hands"

In early 2026, Jordanian company RightNow, which specializes in developing GPU infrastructure and AI solutions for it, released the OpenFang operating system. It is written in Rust and distributed under a dual license: MIT and Apache 2.0. As specialists write, frameworks for working with AI systems are not autonomous enough. Agents have to be prodded with a stick to make them complete tasks, and they often sit idle waiting for user confirmation. That is why RightNow decided to develop an operating system that will help agents become more independent.

The key feature of OpenFang is the autonomous AI modules called "hands". Each such component is responsible for a strictly defined range of tasks — for example, data analytics, collecting information from open sources, forecasting, conducting in-depth research, or working with the web. The "hands" come with a HAND.toml configuration manifest, a multi-step system prompt, a description of the AI skill in SKILL.md format, as well as a set of dashboard metrics. At the same time, users can implement their own autonomous AI agent — the developers have shared relevant instructions for this.

OpenFang supports work with more than 20 different language models, and also offers dozens of ready-made templates for AI agents. For example, for OWASP vulnerability security auditing, you can choose one of the DeepSeek models, while Llama 3.1 8B Instant and Gemini 2.5 Flash are suitable for code monitoring and review respectively.

Overall, developers compared OpenFang's capabilities with the metrics of OpenClaw, LangGraph, CrewAI, and AutoGen, as well as ZeroClaw. In terms of cold start speed, the new agent OS showed one of the best results: just 180 ms — faster than most competitors (the only exception was the lightweight ZeroClaw, which took about 10 ms). OpenFang also proved to have a significant advantage in the number of built-in security mechanisms. The OS offers an isolated WASM sandbox, Ed25519-signed agent manifests for identifying AI systems, as well as a mechanism for erasing secrets from memory after use. The two-way authentication protocol OpenFang Wire Protocol based on the HMAC-SHA256 cryptographic algorithm has also been developed for the OS.

The entire OpenFang is compiled into a single binary file weighing about 32 MB and is ready to use out of the box — installation and basic commands are described in the repository. If you want to learn more about the OS's capabilities, the official documentation is a great starting point — it includes the solution architecture, guides for writing AI skills, API references, and CLI command manuals. It is worth noting that the project is still young, and developers warn about possible bugs and critical changes. But just a few months after its release, OpenFang has gained more than 17 thousand stars on GitHub and received over 2 thousand forks.

Step Towards Full Autonomy

Linux kernel schedulers use universal CPU time allocation strategies and often do not take into account the needs of specific applications. Researchers from Shanghai and Connecticut universities identified this as a problem and last year presented SchedCP —  an experimental MCP server that allows managing Linux schedulers and automatically optimizing processes using AI. According to the authors, their project is a step towards fully autonomous agent operating systems.

SchedCP analyzes the load on the fly, builds its profile, and switches between different schedulers. The system works as an intermediate intelligent layer: agents determine which parameters need optimization, and the OS chooses how to do it. An important role in this architecture is played by the autotune component. The utility collects information and runs the task using the default scheduler. Then the AI agent evaluates performance metrics, builds the load profile, and sequentially tests different schedulers to determine the best option for the given configuration.

To demonstrate SchedCP's capabilities, researchers presented the multi-agent Sched-Agent system, which, according to their words, «mimics the work of a team of experts». The system uses four agents: the first analyzes the load to create an optimization strategy, the second determines tasks for the scheduler, the third handles their implementation, and the last analyzes the results to improve the system's operation.

Instructions for setting up SchedCP, including documentation for complex scenarios — for example, creating custom BPF schedulers — are available in the repository. All published code and accompanying materials are distributed under the MIT license.

DIY Multi-Interface

The goal of the Lukan project is to create an agentic workstation. It provides an AI layer on top of the user's operating system for working with various neural networks, where each model runs in a separate session using tmux or PTY. The project was introduced by an AI systems developer and professor at the University of Alcalá in Spain.

As the author himself explains, Lukan grew out of a personal pain point. Neural networks have become an integral part of his work: he used them for writing code, managing tasks, supporting releases, and a host of other processes. Over time, the number of AI tools he used grew so large that coordinating their work became difficult. This is how the Spanish engineer came up with the idea of creating an OS in which he could control a fleet of agents.

Workflows in Lukan are organized based on a directed acyclic graph (DAG) — you can break a task into stages and assign a specific AI agent to complete each of them. For example, the hypothetical Qwen 3.5 model analyzes context — the user's project, tools, source code, configuration, and history. Then, based on this data, it distributes tasks among AI systems: which commands need to be executed, which APIs to call, and so on.  Meanwhile, GPT-5.3-Codex reports the results to the user in the chat.

The environment supports both open and proprietary models from different providers. Commands can be sent to AI agents via messengers or email. The project author has also considered some security nuances — the solution can automatically detect confidential files based on specified patterns and block AI agents' access to them. For example, you can prohibit neural networks from interacting with credentials, private keys, and service files such as .env, .ssh, and .aws. All information about installing and working with the environment is located in the project's repository. There you can also find instructions for scheduling background tasks in Lukan, as well as a description of the solution's architecture.

More than just an “operating system”

Finally, a slightly unusual and even niche project. Waycore is not just an agent OS, but a modular field computer designed for wilderness survival (and everyday use). The device was designed in 2026 by a Canadian developer who was inspired by the ease of customization and modular structure of the Flipper Zero gadget.

The built-in AI assistant in Waycore works offline and helps identify surrounding objects from images, for example, plants and mushrooms. Also, via the AI agent you can access a local knowledge base with information about survival, navigation, first aid, knot tying, and identification of edible plants. The author filled the knowledge base from specialized resources and manuals. However, only one AI model for text tasks and object recognition can run in Waycore at a time — by default, these are Phi-3 Mini 4K and MobileNetV3. You can also load your own, as Waycore is compatible with the GGUF and TensorFlow Lite formats. The platform supports multiple communication modes that can be switched depending on signal strength and environment: Meshtastic (LoRa mesh), LTE, Wi-Fi, TAK / Cursor on Target (COT) over IP [there is also a beacon / SOS mode]. Some help establish communication in remote areas or even maintain it when there is no internet, while others help determine location and make emergency calls. Support for VoIP is planned to be added in the future when an LTE or Wi-Fi connection is available.

As for the hardware, Waycore is built on a single-board Linux device and an ESP32-S3 microcontroller. It has various sensors: gyroscope, barometer, thermometer, compass, light sensor. Solar panels for charging the built-in battery are located on the back of the device. If desired, the field computer's capabilities can be expanded using humidity or water quality sensors, laser rangefinders, RFID/NFC readers, and so on. All information about the device, including developer guides — for example, how to write applications for the device — is available in the repository for Waycore. There are also guides on usage, setup, debugging, an architecture overview, and instructions for loading AI agents. The source code is distributed under the MIT license.