AnswersPublished23 sources
DeepSeek Harness: A Plugin-First Runtime for Custom AI Agents
DeepSeek Harness (dsh) is an MIT licensed agent runtime in developer preview, not an AI model. Harness ships with Standard, Code, Minimal, and Creator runtime modes, while its append only session log makes runs inspectable through resume, fork, search, and replay.
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- DeepSeek Harness (dsh) is an MIT licensed agent runtime in developer preview, not an AI model.
- Harness ships with Standard, Code, Minimal, and Creator runtime modes, while its append only session log makes runs inspectable through resume, fork, search, and replay.
- Start with the npm CLI, make configuration explicit, and validate plugins both alone and in their full composed runtime before promoting changes.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is DeepSeek Harness, the MIT-licensed developer-preview agent runtime released in September 2026, and how do its “everything is a plugi. Article summary: DeepSeek Harness (`dsh`) is an open-source, MIT-licensed agent runtime in developer preview—not a model—designed to build agents by composing replaceable runtime capabilities rather than modifying a fixed core. The provi. Topic tags: general, education, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, cha
DeepSeek Harness, also called dsh, is an open-source agent runtime for developers building and operating AI agents. It is MIT-licensed and remains in developer preview. Rather than supplying a fixed agent product or a new foundation model, Harness provides an environment in which the parts of an agent can be selected and assembled. Available reporting dates the v0.1 release to August 13, 2026—not September. 1
2
The central idea: no privileged core
Harness takes “everything is a plugin” literally. Model adapters, tools, skills, sessions, sandboxes, storage, scheduling, the agent loop, and the UI can be swapped or recomposed through configuration. The point is to extend a runtime by mounting a plugin beside existing components instead of patching a core implementation. 2
21
That matters when an agent needs a different model provider, a narrower tool set, a custom approval policy, or a distinct execution environment. Those choices can be expressed as runtime composition rather than maintained as a long-lived fork.
What Cordis does
The system is built on Cordis, the vendored plugin framework beneath Harness. A plugin contributes services, typed events, and reversible effects to a shared context. When a plugin unloads, its registrations can unwind as well. 21
22
Cordis also supports dependency-aware composition. A callback registered against required services is unloaded and run again if a required service changes; ctx.plugin loads a plugin using configuration validated against its declared schema. 19
For extension authors, this creates clear seams for customization. A tool plugin can register on ctx.tools, while a pre-execution policy hook can return a typed allow-or-deny decision. The same hook point can support permission, sandbox, and plan-mode policies. 17
Four supplied runtime modes
Harness includes four starting presets for different development tasks: 2
- Standard provides the full coding-agent workflow and tool set.
- Code exposes tools through a Code Mode SDK so model-generated TypeScript can coordinate multistep work.
- Minimal provides a persistent shell and
str_replace_editor, making it useful for constrained setups and benchmarking. - Creator adds runtime inspection and in-memory plugin experimentation for developers creating new presets.
These presets are starting compositions, not permanent boundaries. The plugin architecture is intended to let teams replace or add capabilities as their requirements change. 2
21
Installation and first-run workflow
Harness is distributed as the @deepseek-ai/dsh npm package and CLI. A documented quick start is:
npx @deepseek-ai/dsh web
Developers can also clone the source repository when they need to run from source or work directly with the codebase. 1
30
In the Web UI, configure a model credential under Settings → Models, then select a workspace before composing tasks. The dsh process uses the directory from which it was invoked as its default filesystem location, but a fresh UI has no selected workspace until one is added. 25
Observability is built into the session model
Harness records what the model sees in an append-only session log: system prompts, reasoning, tool calls and results, subagent scheduling, and context injections. Its Trajectory view can inspect these records by source, while resume, fork, search, and replay all use the same event stream. 2
This design makes the log more than a debugging afterthought. It is the record used to examine a run, revisit a prior branch, and reproduce the sequence of runtime events that shaped an agent’s behavior.
Developer preview means compatibility needs discipline
Harness is explicitly a developer preview, so teams should not assume that plugins and APIs carry a stable long-term compatibility promise. 2
6
A sensible operating approach is to pin the version in use, review release changes, test upgrades in an isolated profile or workspace, and replay representative sessions before wider deployment. That is a practical risk-control recommendation based on preview status—not a documented guarantee that every upgrade will be backward-compatible.
Plugin validation should happen at two levels:
- Validate the plugin itself. Confirm its configuration schema, required services, failure behavior, and tool-policy logic.
- Validate the composed preset. Test the plugin with the real model adapter, tools, sandbox, permissions, session store, and other plugins it will run alongside.
Harness configuration is designed to help here. A cordis.yml entry can include a config block, and a plugin exposes a schema that is validated before its apply function runs; invalid configuration fails the load rather than starting the plugin half-configured. 27
28 The generated configuration catalog is checked with
pnpm run verify-config-catalog. 18
For distribution, the documentation supports packaging an extension as an installable bundle and adding it to a profile with dsh plugin add. Keeping bundles and configuration explicit makes a composed runtime easier to reproduce, audit, test, and upgrade. 24
Bottom line
DeepSeek Harness is best understood as infrastructure for agent builders: a runtime where the agent’s model connection, tools, policies, execution environment, loop, state, and interface are composition choices. Cordis supplies the plugin lifecycle; runtime modes provide usable defaults; and the append-only session log provides a detailed operational record. Its flexibility is the attraction, while its developer-preview status makes deliberate testing and upgrade practices essential. 2
21
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "DeepSeek Harness: A Plugin-First Runtime for Custom AI Agents"?
DeepSeek Harness (dsh) is an MIT licensed agent runtime in developer preview, not an AI model.
What are the key points to validate first?
DeepSeek Harness (dsh) is an MIT licensed agent runtime in developer preview, not an AI model. Harness ships with Standard, Code, Minimal, and Creator runtime modes, while its append only session log makes runs inspectable through resume, fork, search, and replay.
What should I do next in practice?
Start with the npm CLI, make configuration explicit, and validate plugins both alone and in their full composed runtime before promoting changes.