openJiuwen ships RSI-native AgentOS, letting agents hot-plug features they write themselves

openJiuwen has released a runtime it calls RSI-native AgentOS, built around a simple proposition: an agent working on a task should be able to give itself a new capability mid-flight. A QbitAI article published under an authorised reprint uses the company's WorkSwarm app to show the loop, in which a user describes what they want and the agent generates an extension package that takes effect immediately without touching the application's original code.

The framing is a shift from developer-friendly to RSI-friendly. RSI stands for recursive self-improvement, which openJiuwen lists as one of eight characteristics it associates with its HyperAgent concept. An extension mechanism, in this view, should not only make it easy for developers to wire in capabilities, but also let the agent act on task feedback, change itself, install the change and test it before the next round.

The architecture has three layers. A web client handles the interface and user actions, a gateway admits and routes requests, and the agent layer runs sessions through AgentServer while a Harness organises task execution and tool calls. All three expose extension entries, which resolve into five hook points: Client, Gateway, AgentServer, Swarm and Agent.

Changes ship as extension packages. A manifest.json declares the package identity, version, dependencies and code entry point, and can specify whether a UI component belongs in the input area, sidebar or settings page. On install the runtime validates the manifest and stores the package by ID and version; on load, the gateway and AgentServer each pick the packages they run, register their handlers on an event bus and fire them when a button is clicked, a request arrives or the agent begins assembling a prompt.

Handlers reach the system through operations the article calls Ops: reading session history and context usage, registering tools for the agent, querying team task progress, updating the workbench interface, making HTTP requests through the gateway, and reading or writing files and the extension's own state within authorised scope. Interface changes and execution changes sit on the same interface rather than being limited to reskinning.

The examples are concrete. Told to add a desktop pet to WorkSwarm, the agent generates an extension package and installs it as an application-form system plugin; the workbench collapses to a clickable pet on the desktop, and a double click returns to the workbench with the conversation intact. In another, an extension adds a status bar above the input box that reports token usage as weather, and the agent tightens the layout after the user says it takes up too much room.

Version management and rollback are what make the loop safe to use. Packages are stored by version, and when a new build breaks a page the user can restore the configuration from a savepoint taken before the failure, then keep debugging. The article is explicit that being able to roll back is what gives people room to try and retry. New sessions can also run on a chosen Harness, with Claude Code, Codex and openJiuwen among the options.

WorkSwarm is open source at github.com/openJiuwen-ai/jiuwenswarm, with extension documentation tied to the 0.2.9.beta3 release. The open question is where self-modification stops being a demo: extensions can register tools, adjust prompts and route execution into third-party sandboxes, so how permissions are narrowed and how enterprise deployments are audited will decide whether RSI-native runtimes reach real team workflows.