DeepSeek's new agent harness makes every component a plugin
DeepSeek has released an early open source version of an agent harness called DeepSeek Harness. A harness is the middleware that sits between a user and an AI model, handling prompts, context management, tool orchestration, the agent loop, state, error handling, safety and permissions; the term came into wide use this year and now covers products such as Claude Code (for Anthropic's Claude models), Codex (for OpenAI's GPT models), and others including Aider, Cline, Goose, OpenCode, OpenHands and Pi. Harnesses have become a key competitive battleground because as models are commoditized, the harness increasingly determines the user interface and locks developers into the tooling they have configured, and because studies have found the same model produces different results depending on the harness that runs it. As an example, the Pi coding agent uses a minimal system prompt of about 200 tokens, while Claude Code used a system prompt of around 10,000 tokens until last month, when Anthropic trimmed it by about 80 percent.
DeepSeek Harness is built on DeepSeek's Cordis meta-framework around the idea that everything, models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration and UI, is a plugin that can be mixed, matched, replaced or extended. According to the DeepSeek Harness website, Cordis services and events let plugins work together, and developers can select, swap or extend any capability through configuration without touching the harness's source code. A DeepSeek paper on Cordis, credited to researchers Yifan Shi, Wei Zhang and Tianyi Cui, describes the framework as supporting dynamic composability: plugins can be added and removed on the fly without breaking the application. The paper distinguishes temporal composability, removing a component and reverting its effects, from spatial composability, letting components manage dependencies on other components, and argues both are necessary in a system that can be modified continuously with little or no human oversight. As a point of comparison, the paper cites Visual Studio Code's extension system: VS Code runs all extensions in a shared extension host process that must be restarted to remove an activated extension, and while VS Code lets extensions declare dependencies on each other, that mechanism is rarely used; DeepSeek Harness supports plugin dependencies directly.
DeepSeek Harness also records everything a model sees, system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection, in an append-only session log. A Trajectory view lets developers inspect these records by source, and resume, fork, search and replay all operate on the same event stream. That matters because access to a model's intermediate reasoning, the chain-of-thought traces DeepSeek's R1 model popularized last year, helps assess whether a model is reasoning well, whether its answers are accurate, and how additional thinking changes output. Anthropic exposes some of this via extended or adaptive thinking depending on the model, but has increasingly summarized rather than shown raw chain-of-thought traces; it says the text shown in a thinking block is a summary of Claude's reasoning, and has implemented classifiers for detection of chain-of-thought elicitation used to construct reasoning training data, a move the article says appears partly aimed at preventing rivals from distilling its models by training on its raw reasoning traces. Raw thinking access from Anthropic requires contacting its sales team. OpenAI has similarly kept chain-of-thought hidden outside its open source models, saying when it introduced its o1 reasoning model two years ago that after weighing multiple factors including user experience, competitive advantage, and the option to pursue the chain of thought monitoring, it had decided not to show the raw chains of thought to users. By contrast, DeepSeek's newly released DeepSeek-V4-Pro and V4-Flash models have thinking mode enabled by default in the API.
Armin Ronacher, co-founder of AI company Earendil, which now steers development of the Pi coding agent, praised the release in a social media post: he does not think DeepSeek Harness is perfect, but said it is the first time he has looked at something new in the space and felt inspired to revisit some of his own choices, adding that he loves that part about open source.
Key facts
- DeepSeek released an early open source version of DeepSeek Harness, an agent harness built on its Cordis meta-framework, in which every component, models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration and UI, is a swappable plugin.
- Cordis is built for dynamic composability: temporal composability lets a plugin be removed with its effects reverted, and spatial composability lets plugins manage dependencies on each other, unlike VS Code's extension host, which the paper's authors say must be restarted to remove an extension.
- A DeepSeek paper on Cordis is credited to researchers Yifan Shi, Wei Zhang and Tianyi Cui.
- The harness logs every system prompt, reasoning step, tool call and subagent action in an append-only session log, viewable in a Trajectory view that supports resume, fork, search and replay.
- Armin Ronacher, co-founder of Earendil, which now steers the Pi coding agent, called DeepSeek Harness the first thing in the space to make him feel inspired to revisit some of his own design choices.
Why it matters
As models become commoditized and harder to tell apart on benchmarks, the harness wrapped around a model increasingly decides how it performs and how developers experience it: the same model produces different results depending on the harness that runs it, and once developers configure their tooling, switching away becomes a source of inertia. DeepSeek Harness stands out for its everything-is-a-plugin design and because it signals Chinese AI labs competing beyond model benchmarks and pricing, into developer tooling itself.
Who it affects
Developers building or running AI agents, and the teams behind competing harnesses such as Claude Code, Codex, Aider, Cline, Goose, OpenCode, OpenHands and Pi. The release specifically drew a public reaction from Armin Ronacher, co-founder of Earendil, which steers Pi's development, who said it made him reconsider his own harness's design choices.
How to use it
DeepSeek Harness is available now as an early open source release built on DeepSeek's Cordis meta-framework. Every capability, models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the UI, is implemented as a plugin that can be selected, swapped or extended through configuration rather than by modifying the harness's source code, and it supports declared dependencies between plugins. It also ships a Trajectory view over an append-only session log, letting developers resume, fork, search and replay a model's full history of prompts, reasoning, tool calls and subagent activity.
How solid is it
The design is documented in a DeepSeek paper credited to researchers Yifan Shi, Wei Zhang and Tianyi Cui, which lays out the temporal and spatial composability concepts behind Cordis and contrasts them with Visual Studio Code's extension host, an established, widely used system. The release has drawn a positive independent reaction from Armin Ronacher of Earendil. It remains an early version, and the broader claim that harness choice significantly affects model performance and cost is attributed only to unnamed various studies, not to a specific benchmark cited in the article.
Risks and caveats
DeepSeek Harness is an early release, so its stability and completeness at scale are unproven. The term harness itself is used loosely across the industry, sometimes meaning just the agent loop and tools, sometimes a much broader set of concerns, which makes direct comparisons between products imprecise. The article's suggestion that Anthropic's summarizing of chain-of-thought is partly meant to prevent distillation is presented as something that appears to be the case, not a claim Anthropic itself has confirmed.
“I don't think the DeepSeek Harness is perfect but this is for sure the first time I have been looking at something new in the space and felt quite inspired to revisit some of our choices. I love that part about Open Source a lot!”
— Armin Ronacher, co-founder of AI company Earendil, which now steers development of the Pi coding agent