DeepSeek has once again captured widespread attention in the developer community, this time focusing on its newly released open-source intelligent agent project, DeepSeek Harness (DSH), which is now available for public testing globally.

On August 13, the DeepSeek Harness v0.1 developer preview was officially opened for testing to developers worldwide, with its source code simultaneously released under the MIT license.

The project adopts a core design philosophy of "everything is a plugin," where all components—including models, tools, sandboxes, UIs, and even the Agent Loop itself—are fully modularized, allowing for free replacement and flexible combination.

Following its release, the project quickly sparked heated discussions on platforms like X, with several early access users publishing in-depth reviews that attracted significant attention. The most widely circulated comment came from early tester Jiayuan (JY) Zhang.

On X, he noted that DSH already exhibits the rudiments of a "self-evolving software." The agent can inspect its own runtime, write and mount plugins on the fly, and directly call upon this newly acquired capability in subsequent tasks.

While he acknowledged that the dynamically generated plugins currently exist only in memory and disappear after a restart, making the functionality still experimental, the potential direction hinted at by this feature has sparked strong interest in the developer community.

Core logic of the architecture: "Everything is a plugin"

DeepSeek Harness is built on a plugin system called Cordis.

The Cordis meta-framework is responsible solely for loading, unloading, and managing dependencies among plugins. All specific components of the Agent Harness—including models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UIs—exist as independent plugins. They collaborate through Cordis services and events, and can be freely combined at the configuration layer.

The key value of this architecture is that developers can independently select, replace, or extend any capability without modifying the DSH source code itself. DeepSeek officially defines this as the project's most important design principle.

Early tester Jiayuan (JY) Zhang compared this design to a "Lego car toy": the Coding Agent provided by DeepSeek is just a pre-built set they assembled. Developers can replace any component and ultimately create something entirely unrelated to a car.

He believes this leaves substantial room for exploration within the future community ecosystem.

Self-evolving characteristic: potential and limitations coexist

Jiayuan (JY) Zhang elaborated on the longer-term possibilities suggested by DSH's creation mode in his review.

He described a scenario: when an agent discovers it lacks a certain capability during a task, it can autonomously complete the development, install the component, and directly invoke it, all without human intervention. He wrote:

"Suppose a feature doesn't exist now. You chat with the Agent for a bit, and the feature is built and ready to use immediately."

However, he also pointed out that this feature is still experimental: dynamically generated plugins exist only in memory and vanish after a restart, and they cannot yet automatically settle into permanent plugins.

In response, another X user, 宝玉@dotey, expressed a notably cautious view. He wrote:

"Software self-evolution might be a false proposition, only leading to greater chaos. Plugins are either disposable after one use or require design, validation, and maintenance—something the current model capabilities cannot achieve through 'self-evolution.'"

He also cited the case of OpenClaw, arguing that the "self-evolving" Skills there had already set a negative precedent, adding, "It's more reliable to wait for models to self-learn and self-evolve."

User experience shortcomings: the gap between geek tools and the mass market

Setting aside conceptual discussions, user 宝玉 offered more direct product feedback from a typical user's perspective.

He pointed out that DSH's current startup method is overly geeky—relying on the Node.js development toolchain, making it nearly impossible for ordinary users to run.

On the experience side, he acknowledged the processing speed but noted issues like rapid text flickering in the chain-of-thought display due to the model's speed, and a lack of output preview capabilities similar to the multi-tab panel on the right side of Codex. He also suggested that developer-oriented features like Trajectory and session logs should not occupy such prominent positions.

His core argument was that the prosperity of a plugin ecosystem depends on a large base of ordinary users, and accumulating such users hinges on an out-of-the-box experience quality. User 宝玉 concluded:

"Developers will only bother to create plugins if your Agent Harness has a massive user base. Otherwise, few developers will stick around, and once the novelty fades, it will be abandoned."