DeepSeek Harness is best understood as infrastructure for building and running agents, not as an AI model in its own right. The developer preview, released under the MIT license through the dsh command-line tool, is built around a simple equation: Model + Harness = Agent. The model supplies reasoning; the harness supplies the operational layer that connects that reasoning to files, terminals, tools, sessions, and execution. 69

That distinction is the main finding from the August 21 hands-on review. DeepSeek’s bet is less about shipping a polished standalone coding assistant and more about making the machinery behind an agent configurable and extensible.

What “Model + Harness = Agent” means

A language model can plan and generate instructions, but it cannot independently inspect a local project, edit files, run commands, maintain an execution history, or coordinate repeated tool calls. The harness provides those capabilities around the model.

In DeepSeek Harness, the model adapter, tool registry, session log, and agent loop are themselves replaceable plugins. The architecture is built on Cordis, whose plugins contribute services, typed events, and reversible effects to a shared runtime context. 17

That separation has two important consequences:

• The model is not the whole agent. It makes decisions and produces reasoning or actions.
• The harness is not tied conceptually to one model. Its runtime can be configured around different model adapters and capability sets, rather than treating the model and agent application as a single inseparable product. 16

This is why calling DeepSeek Harness a “new AI model” misses the point. It is an execution runtime for turning a reasoning model into a system that can act.

How to start the developer preview

The reviewed quick-start path used the following command:

npx @deepseek-ai/dsh web

The local web interface was served at localhost:3080; another hands-on report describes the equivalent local address as 127.0.0.1:3080. 64

The review also reported entering a DeepSeek API key and encountering an interface branded “Exploring the Uncharted.” Those are observations from the hands-on experience rather than claims that every installation will present the same interface or configuration flow. 7

The practical takeaway is that the preview is aimed at developers who are comfortable configuring a local runtime and supplying their own model access—not users looking for a finished, zero-setup coding assistant.

Four work modes, four plugin compositions

DeepSeek Harness does not treat its modes as four unrelated products. They are preset runtime compositions: each loads a different default collection of plugins for a particular kind of work. 1823

Standard: the full agent experience

Standard is the general-purpose preset. DeepSeek describes it as a full coding agent with file editing, shell access, file and web search, skills, planning, goals, subagents, and workflows. 20

It is the closest fit for everyday agentic development and task execution, although the preview’s rough edges make it more suitable for experimentation than for users who simply want a polished replacement for an established coding agent.

PTC: programmatic tool calling

PTC, or Programmatic Tool Calling, is designed for workflows in which the model writes code to coordinate multiple tool calls. Instead of requesting every tool operation separately, the model can use a generated program to orchestrate a sequence of actions. 20

That approach is intended for stronger models and advanced users handling multi-step tasks. It can make tool-heavy workflows more structured, but it also increases the importance of inspecting what the generated code will do.

Minimal: a controlled test environment

Minimal strips the runtime down to a shell tool and a file editor. DeepSeek positions it as a way to benchmark or study models in a deliberately constrained environment, with less assistance from the surrounding harness. 20

This mode is useful when the goal is evaluation rather than convenience. It helps separate a model’s ability from the capabilities supplied by a full agent stack.

Creation: a mode for building modes

Creation is aimed at developers modifying or extending the harness. It supports runtime inspection, in-memory Cordis plugin experiments, and the composition of custom presets. 20

A documented Creation-mode plugin also adds Cordis self-inspection, dynamic-plugin experimentation, and guidance for authoring presets. 22 That makes Creation the clearest expression of DeepSeek’s larger ambition: the system is intended not only to run agents, but also to let developers reshape the runtime those agents use.

Cordis and the “everything is a plugin” architecture

The most distinctive part of DeepSeek Harness is its architectural philosophy. Rather than reserving a large, fixed core for the model, tools, history, and control loop, it exposes those pieces through the Cordis plugin system.

The official architecture documentation specifically identifies the following as replaceable plugins:

• model adapters;
• tool registries;
• session logs; and
• the agent loop itself. 17

The hands-on reporting describes a broader vision spanning skills, sessions, sandboxes, storage, scheduling, and the user interface. Some of those broader examples are not independently established by the supplied primary-source excerpt, so they are better treated as the project’s reported direction than as equally verified implementation details. 117

This design also explains why the four modes can feel fundamentally different without requiring four separate applications. A profile can compose ordered plugin layers at startup, allowing a runtime to be assembled for ordinary coding, controlled evaluation, programmatic tool use, or plugin development. 1723

Observable execution instead of a black box

Agent systems are difficult to debug when they expose only a final answer. DeepSeek Harness places greater emphasis on the execution path: sessions and trajectories are part of the runtime architecture, and the session-log component is replaceable like other plugins. 17

A separate hands-on report described the trajectory view as an append-only JSONL event stream on disk, which made it possible to diagnose a configuration problem quickly. 4 The supplied official architecture excerpt confirms session logs as plugins but does not, by itself, verify every implementation detail of the append-only format.

For developers, this kind of traceability may matter as much as raw model quality. It can help answer questions such as which tool was called, what the agent attempted, and where a long task began to drift.

What the early tests reveal about performance

The available tests point to real capability, but they do not establish a universal performance benchmark. In one August 21 session using DeepSeek V4 Flash 0731, the agent spent 46 minutes and 109 steps building a plugin, consuming 8.9 million input tokens. 1 Another hands-on test reported roughly 20 million tokens over about 35 minutes while building an International Space Station tracker. 2

Those results support a cautious conclusion: long autonomous tasks can be slow and expensive in the developer preview, with runtime depending on the model, task, caching, and workload. They should not be read as a fixed speed limit for every DeepSeek Harness session.

The same evidence also clarifies the product’s current trade-off. DeepSeek is offering unusually configurable infrastructure before it has delivered a fully refined user experience. The project has been described as promising infrastructure and a possible foundation for a contained engineering pilot, but not yet as a production-ready control plane. 56

What the review ultimately revealed

The strongest positioning is not “DeepSeek has built the best coding-agent interface.” It is that DeepSeek has open-sourced a broad agent runtime in which models, tools, execution loops, and other capabilities can be composed and replaced.

That makes DeepSeek Harness particularly interesting to:

• developers building custom agents;
• researchers evaluating models with different levels of harness assistance;
• teams that need inspectable execution histories; and
• open-source contributors experimenting with plugins and runtime composition.

It is less compelling today for someone who only wants a polished coding agent with predictable latency and minimal setup. The v0.1 shell is still early, and the long-task reports show that capability does not automatically translate into speed or production readiness. 56

In short, “Model + Harness = Agent” is both a product explanation and a strategic statement. DeepSeek is separating intelligence from the machinery that lets intelligence act—and betting that an extensible plugin runtime can become more valuable than a single fixed agent application.

DeepSeek Harness is best understood as infrastructure for building and running agents, not as an AI model in its own right. The developer preview, released under the MIT license through the dsh command-line tool, is built around a simple equation: Model + Harness = Agent. The model supplies reasoning; the harness supplies the operational layer that connects that reasoning to files, terminals, tools, sessions, and execution. 69

That distinction is the main finding from the August 21 hands-on review. DeepSeek’s bet is less about shipping a polished standalone coding assistant and more about making the machinery behind an agent configurable and extensible.

What “Model + Harness = Agent” means

A language model can plan and generate instructions, but it cannot independently inspect a local project, edit files, run commands, maintain an execution history, or coordinate repeated tool calls. The harness provides those capabilities around the model.

In DeepSeek Harness, the model adapter, tool registry, session log, and agent loop are themselves replaceable plugins. The architecture is built on Cordis, whose plugins contribute services, typed events, and reversible effects to a shared runtime context. 17

That separation has two important consequences:

• The model is not the whole agent. It makes decisions and produces reasoning or actions.
• The harness is not tied conceptually to one model. Its runtime can be configured around different model adapters and capability sets, rather than treating the model and agent application as a single inseparable product. 16

This is why calling DeepSeek Harness a “new AI model” misses the point. It is an execution runtime for turning a reasoning model into a system that can act.

How to start the developer preview

The reviewed quick-start path used the following command:

npx @deepseek-ai/dsh web

The local web interface was served at localhost:3080; another hands-on report describes the equivalent local address as 127.0.0.1:3080. 64

The review also reported entering a DeepSeek API key and encountering an interface branded “Exploring the Uncharted.” Those are observations from the hands-on experience rather than claims that every installation will present the same interface or configuration flow. 7

The practical takeaway is that the preview is aimed at developers who are comfortable configuring a local runtime and supplying their own model access—not users looking for a finished, zero-setup coding assistant.

Four work modes, four plugin compositions

DeepSeek Harness does not treat its modes as four unrelated products. They are preset runtime compositions: each loads a different default collection of plugins for a particular kind of work. 1823

Standard: the full agent experience

Standard is the general-purpose preset. DeepSeek describes it as a full coding agent with file editing, shell access, file and web search, skills, planning, goals, subagents, and workflows. 20

It is the closest fit for everyday agentic development and task execution, although the preview’s rough edges make it more suitable for experimentation than for users who simply want a polished replacement for an established coding agent.

PTC: programmatic tool calling

PTC, or Programmatic Tool Calling, is designed for workflows in which the model writes code to coordinate multiple tool calls. Instead of requesting every tool operation separately, the model can use a generated program to orchestrate a sequence of actions. 20

That approach is intended for stronger models and advanced users handling multi-step tasks. It can make tool-heavy workflows more structured, but it also increases the importance of inspecting what the generated code will do.

Minimal: a controlled test environment

Minimal strips the runtime down to a shell tool and a file editor. DeepSeek positions it as a way to benchmark or study models in a deliberately constrained environment, with less assistance from the surrounding harness. 20

This mode is useful when the goal is evaluation rather than convenience. It helps separate a model’s ability from the capabilities supplied by a full agent stack.

Creation: a mode for building modes

Creation is aimed at developers modifying or extending the harness. It supports runtime inspection, in-memory Cordis plugin experiments, and the composition of custom presets. 20

A documented Creation-mode plugin also adds Cordis self-inspection, dynamic-plugin experimentation, and guidance for authoring presets. 22 That makes Creation the clearest expression of DeepSeek’s larger ambition: the system is intended not only to run agents, but also to let developers reshape the runtime those agents use.

Cordis and the “everything is a plugin” architecture

The most distinctive part of DeepSeek Harness is its architectural philosophy. Rather than reserving a large, fixed core for the model, tools, history, and control loop, it exposes those pieces through the Cordis plugin system.

The official architecture documentation specifically identifies the following as replaceable plugins:

• model adapters;
• tool registries;
• session logs; and
• the agent loop itself. 17

The hands-on reporting describes a broader vision spanning skills, sessions, sandboxes, storage, scheduling, and the user interface. Some of those broader examples are not independently established by the supplied primary-source excerpt, so they are better treated as the project’s reported direction than as equally verified implementation details. 117

This design also explains why the four modes can feel fundamentally different without requiring four separate applications. A profile can compose ordered plugin layers at startup, allowing a runtime to be assembled for ordinary coding, controlled evaluation, programmatic tool use, or plugin development. 1723

Observable execution instead of a black box

Agent systems are difficult to debug when they expose only a final answer. DeepSeek Harness places greater emphasis on the execution path: sessions and trajectories are part of the runtime architecture, and the session-log component is replaceable like other plugins. 17

A separate hands-on report described the trajectory view as an append-only JSONL event stream on disk, which made it possible to diagnose a configuration problem quickly. 4 The supplied official architecture excerpt confirms session logs as plugins but does not, by itself, verify every implementation detail of the append-only format.

For developers, this kind of traceability may matter as much as raw model quality. It can help answer questions such as which tool was called, what the agent attempted, and where a long task began to drift.

What the early tests reveal about performance

The available tests point to real capability, but they do not establish a universal performance benchmark. In one August 21 session using DeepSeek V4 Flash 0731, the agent spent 46 minutes and 109 steps building a plugin, consuming 8.9 million input tokens. 1 Another hands-on test reported roughly 20 million tokens over about 35 minutes while building an International Space Station tracker. 2

Those results support a cautious conclusion: long autonomous tasks can be slow and expensive in the developer preview, with runtime depending on the model, task, caching, and workload. They should not be read as a fixed speed limit for every DeepSeek Harness session.

The same evidence also clarifies the product’s current trade-off. DeepSeek is offering unusually configurable infrastructure before it has delivered a fully refined user experience. The project has been described as promising infrastructure and a possible foundation for a contained engineering pilot, but not yet as a production-ready control plane. 56

What the review ultimately revealed

The strongest positioning is not “DeepSeek has built the best coding-agent interface.” It is that DeepSeek has open-sourced a broad agent runtime in which models, tools, execution loops, and other capabilities can be composed and replaced.

That makes DeepSeek Harness particularly interesting to:

• developers building custom agents;
• researchers evaluating models with different levels of harness assistance;
• teams that need inspectable execution histories; and
• open-source contributors experimenting with plugins and runtime composition.

It is less compelling today for someone who only wants a polished coding agent with predictable latency and minimal setup. The v0.1 shell is still early, and the long-task reports show that capability does not automatically translate into speed or production readiness. 56

In short, “Model + Harness = Agent” is both a product explanation and a strategic statement. DeepSeek is separating intelligence from the machinery that lets intelligence act—and betting that an extensible plugin runtime can become more valuable than a single fixed agent application.