DeepSeek has launched an early version of its open source agent harness, built on a distinctive architecture where every component—models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration, and user interfaces—functions as a plugin, according to the China-based AI company. The release signals Chinese AI labs are expanding their competitive focus beyond pricing and model benchmarks. Harnesses have emerged as crucial middleware layers that manage the prompts sent to AI models and the responses they generate, overseeing context management, tool orchestration, the agent loop, state management, error handling, safety, and permissions.
The harness operates on DeepSeek's Cordis meta-framework, which supports what researchers Yifan Shi, Wei Zhang, and Tianyi Cui describe as temporal composability—the ability to add and remove plugins dynamically without requiring application restarts—and spatial composability, which allows components to manage dependencies on other components. A DeepSeek paper explains that Microsoft's Visual Studio Code, by contrast, runs all extensions in a shared process that cannot remove them on the fly without restarting the host. DeepSeek Harness lets developers select, swap, or extend any capability through configuration files without altering the source code itself. The system also records everything the model processes in an append-only session log, capturing system prompts, reasoning steps, tool calls and their results, subagent scheduling, and every context injection, which users can inspect, resume, fork, search, and replay through a Trajectory view.
The harness provides full access to chain of thought traces, which break down prompts into a series of intermediate reasoning steps before producing a final answer. With the newly released DeepSeek-V4-Pro and V4-Flash models, the API enables thinking mode by default, exposing these reasoning steps rather than hiding or summarizing them. Anthropic, by comparison, has been concealing model reasoning by summarizing chain of thought traces and implemented classifiers for "detection of chain-of-thought elicitation used to construct reasoning training data," requiring customers to contact sales personnel to access raw thinking. OpenAI decided not to show raw chains of thought to users two years ago when it introduced its o1 reasoning model, citing user experience, competitive advantage, and the option to pursue chain of thought monitoring as factors.
The DeepSeek researchers argue that temporal and spatial composability are essential in systems where modifications can occur continuously with minimal or no human oversight, serving as a safeguard against forced restarts and crashes when components appear and disappear. Harnesses have become a focal point of competition as models proliferate and commoditize, particularly because they often implement the user interface, creating user inertia once developers configure their tooling and grow accustomed to specific workflows. Different design choices produce significant performance and cost variations—the Pi coding agent relies on a minimal system prompt of roughly 200 tokens, while Claude Code used a system prompt of around 10,000 tokens until Anthropic trimmed it by about 80 percent last month, meaning the same model yields different results depending on the harness. As the open source model ecosystem matures, access to chain of thought traces appears poised to become another avenue for competitive differentiation. Armin Ronacher, co-founder of AI company Earendil, which develops the Pi agent, called DeepSeek Harness the first new project in the space that inspired him to revisit design choices, praising the benefits of open source development. The all-plugin architecture and transparent reasoning access may reshape expectations around flexibility and observability in AI tooling, particularly as developers weigh the trade-offs between proprietary convenience and open configurability.