Long time no see~ I'm Lena.
DeepSeek Harness, or dsh, is an open-source agent runtime as of August 2026, DeepSeek Harness is in developer preview. The short version is: model plus harness equals agent. The model handles reasoning; the harness handles tools, environment, sessions, permissions, sandboxing, and the execution loop. DeepSeek Harness is not simply “DeepSeek’s Claude Code.” It is closer to a modular runtime for composing agents.
What Is DeepSeek Harness?
DeepSeek Harness is an agent harness that runs around a model and decides how the agent operates. The official repository describes dsh as open-source, MIT licensed, in developer preview, and built around one idea: everything is a plugin. A model can answer, but a harness decides what the model can see, which tools it can call, what gets logged, and when approval is required.
npx @deepseek-ai/dsh web
In one sentence: the model thinks; the harness makes that thinking executable inside a controlled environment.
How “Everything Is a Plugin” Works
Cordis as the Plugin Kernel
DeepSeek Harness is powered by Cordis, a meta-framework for composable plugins. I think of Cordis less as “the agent brain” and more as the socket board under the agent. It loads plugins, manages their lifecycle, and lets them register services, events, and reversible effects. Cordis itself does not provide planning, coding, or memory. It gives dsh a way to mount those parts without making them untouchable.
What Can Be Replaced
The architecture notes say the model adapter, tool registry, session log, and agent loop are all plugins, so each part can be replaced through configuration. The same idea extends to Skills, storage, sandbox providers, UI surfaces, filesystem access, and execution policy. There is no privileged core you must patch first.
In a fixed coding agent, you usually accept the packaged loop, tool set, approval model, and UI. In DeepSeek Harness, those pieces become composable infrastructure. Something's a bit unclear here, because plugin freedom also means more responsibility.
What DeepSeek Harness Includes Today
Presets and Runtime Modes
As of the August 2026 developer preview, the runtime modes to verify are Standard, PTC, Minimal, and Creator. I would not treat those names as permanent because the repo warns that compatibility-breaking changes can happen.
Standard is the fuller setup. PTC appears aimed at code-mode style tool batching. Minimal keeps the surface small, usually around persistent shell and editor-style file changes. Creator points toward self-inspecting or plugin-authoring workflows.
Replayable Sessions and Multiple Interfaces
The session model matters. DeepSeek Harness records durable session events, derives model-visible history from that log, and supports patterns such as resume, fork, and replay. If the session is eventized, the runtime can be inspected, continued, or branched instead of becoming a blurry chat transcript.
The current surface includes Web UI, headless execution, Python SDK, JSON-RPC-oriented examples, and protocol-facing paths that teams should verify before wiring into ACP-style editor integrations. It also supports DeepSeek, OpenAI, Anthropic, and custom OpenAI-compatible endpoints.
DeepSeek Harness vs. Coding Agents and Evolution Layers
| System | Positioning | Main responsibility | Experience handling |
|---|---|---|---|
| DeepSeek Harness | Re-composable agent runtime | Tools, sessions, adapters, sandbox, permissions, loop | Records and replays runtime events |
| Claude Code / Codex | Coding agent product experience | Opinionated coding workflow and repo interaction | Product-specific session memory |
| EvoMap | Experience-evolution layer | Structures, validates, and inherits successful agent behavior | Turns experience into reusable assets |
This is not a ranking. A coding agent gives you a working product experience. DeepSeek Harness exposes the runtime pieces behind that experience. EvoMap sits after the run and asks a different question: which successful behaviors should become reusable, auditable, inheritable assets?
I also would not imply official integration here. The relationship in this article is architectural, not partnership-based. For grounding, the DeepSeek Harness repository is the right primary source for dsh, while the Model Context Protocol docs are useful background for the tool-connection layer many agent runtimes now sit beside.
How Agent Experience Can Be Structured After the Runtime
This is an architectural comparison, not a claim of existing product integration.
A harness answers “how does the agent run?” An experience-evolution layer answers “what should survive after the run?” EvoMap frames that second layer through Gene, Capsule, and Event: a way to keep successful behavior from disappearing after one task.
| Asset | Definition | Saved content | Use |
|---|---|---|---|
| Gene | Reusable strategy unit | Reusable strategy template with preconditions, constraints, and validation commands | Let agents inherit a proven tactic |
| Capsule | Verified execution asset | Successful path with context and evidence | Reuse a validated fix or execution pattern with evidence |
| EvolutionEvent | Immutable process record | What happened during execution | Audit, Audit evolution history and validate lineage, and validate lineage |
Harness solves how to run; the evolution layer solves how to reuse what was learned after running.
Who Should Pay Attention—and What Are the Limits?
DeepSeek Harness is worth watching if you build agent infrastructure, multi-model agents, custom tool environments, or sandboxed execution systems.
The limits are real. This is an August 2026 developer preview, not a quiet production contract. APIs and presets may change. The plugin ecosystem is still forming. Plugin architecture also does not automatically make an agent self-improvement. Learning, validation, inheritance, and safe reuse still need a separate design.
FAQ
Does DeepSeek Harness require a DeepSeek model to run?
No. The preview supports DeepSeek, OpenAI, Anthropic, and custom OpenAI-compatible endpoints. Still, test each provider path before assuming identical tool behavior.
Is DeepSeek Harness ready for production use right now?
Treat it as developer preview infrastructure. Production teams should isolate experiments, pin versions, and expect migration work.
Can I use DeepSeek Harness for non-coding agents?
Yes, in principle. Because tools, adapters, storage, sandbox, and UI can be replaced, the runtime is not limited to coding.
How is a DeepSeek Harness plugin different from a typical Agent Skill?
A plugin can change the runtime itself: tools, adapters, events, session behavior, sandbox, or UI. A Skill usually gives reusable instructions inside an existing runtime.
Does a plugin-based harness automatically make an agent self-improving?
No. It makes the agent easier to recompose. Self-improvement needs validated feedback, stored experience, and inheritance rules.
Conclusion
DeepSeek Harness matters because it makes the agent runtime modular. It separates the model from the machinery around it. I’m not ready to call that the whole future of agents, but it does make one thing clearer: after agents can run, the next question is what they keep. That is where EvoMap’s work on reusable agent experience starts to feel relevant.
Previous Posts:
- If you want to see where a modular runtime like dsh fits in the wider agent stack, read OpenHarness and GEP: Where They Sit in the Agent Stack for a closer look at the boundary between harness infrastructure and experience evolution.
- For a broader view of why the agent harness is becoming an infrastructure category of its own, see Harness Engineering: Mem0, LangGraph, and CrewAI.
- To separate runtime execution, tool connection, and agent evolution more clearly, read MCP, CLI, and GEP: Three Layers of the Agent Stack.
- If the difference between DeepSeek Harness plugins, Agent Skills, and reusable experience assets still feels blurry, Agent Skills vs GEP Assets: The Real Difference goes directly into that distinction.
- For the tool-connection layer that sits beside agent runtimes such as DeepSeek Harness, What Is MCP? The AI Tool Connection Standard explains what MCP handles—and what it leaves to the surrounding agent infrastructure.
