DeepSeek released DeepSeek Harness on August 13, 2026, an open-source AI agent execution framework distributed as a developer preview2. The framework treats nearly every component of an agent stack — models, tools, sessions, sandboxes, the execution loop, scheduling, and the user interface — as swappable plugins that can be reconfigured through configuration rather than code changes1.
Harness addresses a layer that sits between a foundation model and a working agent: the runtime machinery for memory, tool orchestration, file handling, and record-keeping. According to the architecture documentation, model adapters, tool registries, session logs, and the agent loop can all be replaced via configuration, and developers can swap existing plugins or add new capabilities without rewriting the Harness core.
At startup, Harness layers multiple bundles in sequence to assemble a named profile. The base bundle contains model connections, tools, and persistence; sandbox and approval policy, credentials, and telemetry are handled by the same layer. Web apps and headless execution are layered on top.
The official distribution ships with four execution modes. Standard mode is a coding agent equipped with file operations, shell, and search, with planning and sub-agent capabilities. Code mode bundles multiple rounds of tool calls together from model-generated code. Minimal mode retains only shell and file editing for evaluating models in a stripped-down environment. Creator mode inspects the running configuration and tests Cordis plugins in memory to create new presets.
Harness records system prompts, inference data, tool calls, results, and sub-agent assignments in an append-only session log. A runtime invariant called context injection requires that content passed to the model be reconstructable from the log. A single model request and its accompanying tool execution are recorded as a step; a process consisting of multiple steps is recorded as a turn. Resumption, branching, search, and replay are derived from the same event sequence.
On data handling, DeepSeek states that input, model output, and session context — including tool execution results, file paths, and API keys — are stored on the user's device by default and are not sent to servers without consent. Data may be sent to external destinations if external models, web tools, MCP services, or external plugins are configured.
The package is licensed under the MIT license, and a method for building from source is published. The official repository warns that breaking changes are coming, and the README states that upcoming changes will break API and configuration compatibility. DeepSeek characterizes the release as a test of the design philosophy rather than a finished product.
An accompanying paper by Yifan Shi, Wei Zhang, and Tianyi Cui of Peking University and DeepSeek-AI defines two composability properties: temporal composability, which involves rolling back side effects when a component is removed, and spatial composability, which involves declaring dependencies between components so they can track changes. The paper notes limitations: Cordis can only roll back to states the framework can track, data already sent externally cannot be automatically restored, and isolating malicious code from the host environment requires a separate sandbox. Interface identifiers can change across versions, and the current implementation handles version constraints via npm's peer dependencies without enforcing that providers adhere to semantic versioning.
With Node.js installed, users can launch the Web UI with the command npx @deepseek-ai/dsh web. Armin Ronacher commented on August 14 that the project "is not perfect but has a freshness that makes him want to reconsider design choices in the field".
ANALYSIS Rather than standardizing inter-agent communication, Harness modularizes the internal runtime of a single agent system. The developer-preview status and explicit breaking-change warnings signal that DeepSeek is soliciting architectural feedback before stabilizing the API.