SI Agent News
DevelopingClaudeOpenAIOpen modelsAgents

openJiuwen's agent platform can write, install and undo its own plugins

A Huawei-backed open-source group built an agent app that adds features to itself on request, with a rollback option. The idea is interesting; the safeguards are unclear.

fiveparts of the stack with extension points

The short version

Why it matters

Software that can extend itself is a different safety question from software that only follows instructions. The rollback feature is the minimum protection. Anyone evaluating agent tools for a team should ask who approves a new plugin, what it can touch and how fast it can be reversed.

An app that adds to itself

openJiuwen, an open-source agent community backed by Huawei, has laid out a design it calls an RSI-native AgentOS. RSI stands for recursive self-improvement, the idea of a system getting better by changing itself. In practice, the design lets an AI agent add features to the very application it runs in, test them in real work, and back them out if they cause trouble.

As pandaily.com reported, the write-up was republished by QbitAI on November 13. openJiuwen describes it as part of its wider Hyper-Agent architecture, and frames the change as a shift from software that is easy for developers to modify toward software that is easy for agents to modify.

How the plugin system is built

The design leaves hooks for new features in five places. The client is the web interface, where plugins can add buttons, panels and settings. The gateway handles login and the routing of requests. The AgentServer deals with sessions, messages and server commands. The swarm layer controls how several agents split up work, pass it along and keep track of progress. The agent layer can tweak prompts and register tools, MCP services (connectors to outside tools) or sub-agents.

Every new feature arrives as a package. Inside is a small file called manifest.json, which names the package, states its version and what it depends on, points to where its code begins, and says where any visible piece belongs on screen. When someone installs a package, the platform validates that file and stores the package under its name and version number.

The gateway and AgentServer then load whichever packages apply to them. Those packages register handlers, which are bits of code that wake up when something happens, such as a user pressing a button, a request arriving, or the agent putting together its standing instructions. Handlers can use operations the components offer, like reading session data or updating the screen.

Packages are hot-pluggable, meaning they can be added to a running system, and doing so does not require editing the application's own code.

What WorkSwarm does with it

WorkSwarm is openJiuwen's agent application, and it uses this mechanism for what the group calls system plugins. You describe the feature you want in plain language. The agent then does the coding, bundles the result and installs it.

The demos were deliberately light. One added a desktop pet mode that shrinks the workbench into a small chat companion. Another, called Token Weather, shows how much of a session's token budget has been used (tokens are the chunks of text a model processes) as a weather display in a status bar.

If a plugin misbehaves, the user can return the plugin setup to an earlier saved state. People can also install plugins by hand, from a folder, from an extension page or with an /extension install command, so the agent is not the only route in.

Enterprise uses and the open code

The same approach is meant to cover business needs. openJiuwen showed a login plugin for the gateway, adapters for third-party sandboxes (walled-off spaces where code runs), and a plugin that lets a session run on a Claude Code, Codex or openJiuwen harness, the surrounding software that drives a model. Larger deployments can run the gateway and AgentServer on separate machines, each loading its own set of extensions.

WorkSwarm's code is public on GitHub as JiuwenSwarm under the Apache-2.0 license, a permissive open-source license. That means outsiders can inspect how the system works, which matters given the questions below.

What is new, and what is not

Plugin systems, version numbers and rollback are old tools in software. The new element, if it works as described, is that the agent plays both author and installer, running the whole loop on its own.

That is also where the risk lies. An agent that can extend itself can widen what it is able to do without anyone asking, and it can install the wrong thing as easily as the right one. The coverage we have does not describe the safeguards: whether a person or an automated check reviews each package, what permissions a package gets, whether installs and rollbacks are logged, or whether packages are sandboxed.

The source also does not give benchmarks or pricing, and it does not say how well rollback holds up in real use. The product is niche, so this is better treated as a pattern to watch than a tool to adopt.

What to watch

Try it

If you evaluate agent tools for a team, ask any vendor three questions: who approves a new plugin, what can it access, and how quickly can you revert it?

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by pandaily.com · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news