DeepSeek has opened a developer preview of DeepSeek Harness, a software framework intended to help developers assemble and inspect AI agents. The company released the source code alongside the preview and warned that the project’s principal interfaces and plugins remain subject to change while testing continues.
The framework treats each major agent function as a plugin. Models, tools, skills, session handling, sandboxes, storage, execution loops, scheduling and the user interface can all be selected or replaced through configuration. DeepSeek says developers can extend these components without editing the core Harness source.
At the centre is Cordis, a kernel responsible for loading and removing plugins and resolving dependencies between them. Services and events allow independently packaged capabilities to communicate. The approach is designed to let builders create different agent environments from a common base instead of maintaining separate, tightly coupled implementations.
DeepSeek Harness also keeps an append-only record of information presented to a model. The log covers system instructions, model reasoning, tool calls and their results, scheduling of subagents, and context added during a session. A Trajectory view lets developers inspect records according to their source. The same event stream supports resuming, branching, searching and replaying sessions.
Four operating modes are included in the preview. Standard mode exposes the complete tool collection. Code mode allows model-generated code to coordinate several rounds of tool use. Minimal mode limits the environment to a shell and file editor, providing a reduced setup for model evaluation. Creator mode is intended for examining a running environment, testing Cordis plugins in memory and combining components into additional modes.
The design addresses a growing layer of AI development that sits around the underlying model. An agent harness must connect a model to external tools, maintain state, manage execution and expose enough information for developers to understand failures. By separating these responsibilities into plugins, DeepSeek is offering a framework in which teams can exchange one implementation without rebuilding the rest of the runtime.
Developers can start the web interface with Node.js and `npx`, or clone the complete source and follow the repository’s setup process. The supplied announcement does not state a production-readiness date, a stable API commitment or service-level guarantees.
Because this is a developer preview, prospective users will need to account for compatibility changes and evaluate security boundaries around tools, sandboxes and stored session data. The release nevertheless provides an open implementation for experimenting with composable agent infrastructure and for examining, replaying and comparing how agents operate.
The preview announcement does not compare Harness with other agent frameworks or provide performance benchmarks, so its practical advantages remain to be established through developer testing.
