DeepSeek has published deepseek-harness, an MIT-licensed runtime for AI agents in which nearly every subsystem is a plugin. According to DeepSeek's official project site, models, tools, skills, sessions, sandboxes, storage, the agent loop, scheduling, and the user interface are all composed from plugins rather than hard-wired. The architecture document goes further and says the model adapter, tool registry, session log, and agent loop are "replaceable from configuration."
Key facts
- Nine subsystems, including the agent loop itself, are plugin-composed according to DeepSeek's own project site.
- Released under the MIT license, the most permissive of the common open-source licenses.
- Built on Cordis, whose companion paper draft is dated August 13, 2026.
- Primary source: github.com/deepseek-ai/deepseek-harness.
Most of the argument about AI agents over the past two years has been an argument about models. DeepSeek's answer this week was not a better model. It was a better substrate for the thing that wraps the model.
Some background helps. When you use a coding agent or a research agent, the model is only one part of what you are talking to. Around it sits what practitioners call a harness: the code that decides what to put in front of the model, which tools it may call, where the conversation is stored, how the sandbox is set up, and when to stop. That harness has quietly become as important as the weights. Earlier this year, swapping only the harness moved DeepSeek by twenty tasks on a coding benchmark with no change to the model at all. The harness is where a lot of the real engineering now lives, and it is usually the least flexible part of the stack.
What DeepSeek did is make that layer modular in a literal, structural way. The runtime is built on Cordis, a plugin framework whose accompanying draft paper is titled "A Programming Paradigm for Spatiotemporal Composability." The paper repo describes two ideas working together: temporal composability, achieved through effects that can be reverted, and spatial composability, achieved through reactive dependencies between parts. In plainer terms, Cordis is trying to let a running program grow and shrink pieces of itself safely, and to have the rest of the system react correctly when it does.
A useful analogy is a stage production. In most agent frameworks the set is nailed down before the curtain goes up, and changing the lighting rig means rebuilding the theater. Cordis is closer to a rigging system where each piece hangs from a standard mount, can be flown in or out mid-performance, and the pieces that depended on it are told about the change instead of quietly breaking.
The claim worth being precise about is the loop one. The DeepSeek Harness source implements the agent loop as a Cordis Service, which means the whole loop driver can be replaced at the runtime boundary. What is not itself a plugin is the ordinary imperative control flow inside that service. The actual while loop in agent.ts is normal code. So "everything is a plugin" is accurate at the level of subsystems and swappable drivers; it is not accurate at the level of individual language constructs. That distinction matters if you are evaluating the project, because the useful property here is that you can bring your own loop, not that the loop is somehow dissolved.
On Lobsters, where the Cordis paper surfaced, the reception was cautiously positive: one commenter read it as a dense but genuinely interesting treatment of how to implement plugin-based software with runtime mounting and unmounting, and another argued that this framework is the real differentiator for DeepSeek's new harness rather than any model change.
Why this matters: if the harness is where the performance is, then the harness is also where the lock-in is. A runtime that lets you replace the model adapter, the tool registry, and the loop independently is a bet that the interesting competition over the next year is at the integration layer, not the weights layer, and that whoever owns a neutral, permissively licensed substrate ends up in a strong position. DeepSeek releasing it under MIT rather than a bespoke license is consistent with that bet, and consistent with how the company has handled its open weights generally.
The honest caveat: several of the claims circulating alongside the launch do not survive checking. There is no primary evidence for the "fastest-starred repository in history" superlative, the upstream GitHub releases page shows no tagged releases at all, and the Cordis paper is a GitHub-hosted draft rather than a published arXiv entry. The engineering claim holds up. The hype around it mostly does not.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (1)
The service boundary is the detail that matters. Replacing the loop driver is valuable, but hot-swapping it safely requires versioned session state and a declared migration path; otherwise a new loop can interpret an old event log differently. The first failure test I’d run is resuming the same session across a loop-plugin change and verifying that tool effects are neither replayed nor skipped.