A
The Agent harness
DeepSeek Harness is an open-source Agent harness developed by DeepSeek AI. Its Cordis-based architecture composes the Agent loop, model adapter, tools, and persistence from plugins and is still in developer preview.
All Agents
Provider · DeepSeek
DeepSeek Harness
Run DeepSeek Harness in a cloud Workspace
Evaluate DeepSeek’s plugin-based harness without first assembling its local runtime. Agent.Space provides the surrounding project Workspace while the live product shows the models and controls currently available.
Use DeepSeek Harness in Agent.Space01 · Harness or model?
Know which layer you are choosing
B
The model underneath
DeepSeek Harness is an Agent runtime and workflow layer, not the DeepSeek model itself. Its plugin-based approach controls how a task is executed; the compatible model is a separate choice. Agent.Space exposes a deliberately narrower bridge while the harness remains in developer preview.
02 · Two ways to run it
Native product or Agent.Space Workspace
Official route
DeepSeek Harness
The official developer-preview route uses DeepSeek’s harness and its plugin-oriented Cordis runtime directly. That path gives developers more control over setup and experimentation, while also making them responsible for an evolving runtime and its compatibility.
Read the official DeepSeek Harness previewAgent.Space route
DeepSeek Harness × Agent.Space
Agent.Space connects to DeepSeek Harness through a controlled automation transport and consumes its committed text responses. The Workspace remains the shared project boundary, while model availability comes from Agent.Space rather than being inferred from the DeepSeek name.
Use DeepSeek Harness in Agent.Space03 · Task fit
Choose it for the work, not the logo
Good fit
- Run focused coding tasks through a composable Agent execution stack.
- Evaluate a plugin-based harness without assembling its runtime locally.
- Compare how the same Workspace task behaves across different harnesses.
Look elsewhere when
- You need a production-stable interface with long-term compatibility guarantees today.
- Your workflow requires image input or user-added MCP servers through the current Agent.Space bridge.
- You need direct control over every Cordis plugin, experimental mode, or upstream runtime detail.
04 · Current product boundary
What Agent.Space supports today
This page describes the current Agent.Space integration, not every feature the upstream vendor offers. The live Workspace controls remain the source of truth.- 01
DeepSeek Harness is still a developer preview. Agent.Space treats it as an experimental option whose behavior and compatibility may change.
- 02
The current controlled automation bridge accepts text input and returns text. It does not accept user-added MCP servers through this integration.
- 03
Only compatible models and controls shown in the live Workspace are supported; upstream plugins and modes are not automatically reproduced.
05 · Workspace workflow
From a bounded task to a reviewable result
Use this integration for controlled evaluation, not as an invisible drop-in replacement for a mature production pipeline. Keep the task small and make the evidence easy to inspect.
- 01
Create an evaluation Workspace
Use a project whose expected outcome and acceptable risk are clear, with an easy path to inspect or discard changes.
- 02
Select DeepSeek Harness
Choose from the compatible models currently exposed by Agent.Space and review the preview status before starting.
- 03
Run a narrow text-based task
Provide explicit inputs and success criteria; avoid depending on unsupported multimodal or MCP-based steps.
- 04
Evaluate the evidence
Review the Session response and Workspace files, record what worked, and move critical follow-up to a stable Agent when needed.
07 · Agent-specific FAQ
Questions to answer before you start
The practical differences that matter when choosing this Agent inside Agent.Space.
Start in one Space
Evaluate DeepSeek Harness inside a controlled Workspace
Start with a narrow text-based task, inspect the evidence, and keep experimental work separate from workflows that require stable guarantees.