Architecture & Plugin Design

Harness is built on the Cordis meta-framework with a radical "everything is a plugin" architecture — models, tools, skills, sessions, sandboxes, storage, agent loops, scheduling, and UI are all implemented as independently swappable plugins that can be freely recomposed . The official positioning formula is "Model + Harness = Agent" — the model handles reasoning while Harness handles execution .

The framework is model-agnostic, runs locally, and ships with features including project management, long-horizon task coordination, multi-agent orchestration, context management, web search, and external skill invocation .

To get started, developers can run a single command: npx @deepseek-ai/dsh web, which launches a local Web UI at http://127.0.0.1:3080 .

Four Execution Modes

Harness provides four preset modes, each loading a different default set of plugins adapted for a specific use case :

• Standard — Full agent mode with complete toolchains, sub-agent dispatch, file editing, shell access, and search. Comparable to Claude Code for everyday agentic work.
• PTC (Programmatic Tool Calling) — Also called Code mode. Tools are exposed via a Code Mode SDK, allowing the model to orchestrate multiple tool calls by generating TypeScript code that chains complex operations.
• Minimal — A bare-bones environment with only a shell tool and a file editor, designed for clean model benchmarking and evaluation without extra tooling.
• Creation/Creator — Custom preset authoring mode. Lets developers inspect the current execution environment, test Cordis plugins in memory, and combine them to build new tailored agent configurations.

All four modes run on the same plugin core — the difference is purely in which plugins are loaded at startup .

Deployment on China's National Supercomputing Internet

On August 14–15, 2026, China's National Supercomputing Internet made DeepSeek-V4-Pro-0813 and DeepSeek Harness v0.1 officially available on its platform — the country's first large-scale integrated computing resource pool with over 100,000 AI accelerator cards . This gives domestic developers access to model deployment, distributed inference, private hosting, and agent development within a single national computing infrastructure, lowering technical barriers for AI adoption in scientific research and industrial applications .

Competitive Positioning Against Claude Code

DeepSeek explicitly positions Harness as a rival to Anthropic's Claude Code and OpenAI Codex, targeting coding and office productivity use cases . The framework team lead, Cui Tianyi, opened global beta testing on August 2, 2026, and attracted dozens of developers before the public launch . The V4-Pro-0813 model also posted competitive agent-focused benchmark scores: Terminal Bench 2.1 at 87.9, DeepSWE at 62.7, Toolathlon-Verified at 74.1, and DSBench-FullStack at 71.1 .

API Price Increases — Reversal of Low-Cost Strategy

On the same day as the Harness launch, DeepSeek announced major API price increases that reverse its previous reputation as the "Token Price Butcher" :

• V4-Pro-0813 is priced at $1.32 per million input tokens and $3.96 per million output tokens — up to 14x higher than the V4 Flash model
• Increases range from 50% to 1,100% above prior prices, depending on the model, token type, and time of use
• A new peak/off-peak pricing structure was introduced, with peak rates rising as much as 12x on cache-hit input tokens — peak hours in China are 9:00–12:00 and 14:00–18:00 Beijing time, with off-peak rates set at half the peak rate
• DeepSeek had signaled a "significant" price hike on August 6, and analysts describe this as the end of the "bargain era" for the company's API services

Why This Matters

The simultaneous releases of Harness and V4-Pro-0813, alongside the pricing pivot, signal a strategic shift at DeepSeek: moving from a pure model provider to an execution-layer infrastructure company. By open-sourcing the agent runtime under MIT and making it available on national supercomputing infrastructure, DeepSeek is betting that the next competitive battleground in AI is not just the model, but the ecosystem of tools, agents, and workflows built around it.

Architecture & Plugin Design

Harness is built on the Cordis meta-framework with a radical "everything is a plugin" architecture — models, tools, skills, sessions, sandboxes, storage, agent loops, scheduling, and UI are all implemented as independently swappable plugins that can be freely recomposed . The official positioning formula is "Model + Harness = Agent" — the model handles reasoning while Harness handles execution .

The framework is model-agnostic, runs locally, and ships with features including project management, long-horizon task coordination, multi-agent orchestration, context management, web search, and external skill invocation .

To get started, developers can run a single command: npx @deepseek-ai/dsh web, which launches a local Web UI at http://127.0.0.1:3080 .

Four Execution Modes

Harness provides four preset modes, each loading a different default set of plugins adapted for a specific use case :

• Standard — Full agent mode with complete toolchains, sub-agent dispatch, file editing, shell access, and search. Comparable to Claude Code for everyday agentic work.
• PTC (Programmatic Tool Calling) — Also called Code mode. Tools are exposed via a Code Mode SDK, allowing the model to orchestrate multiple tool calls by generating TypeScript code that chains complex operations.
• Minimal — A bare-bones environment with only a shell tool and a file editor, designed for clean model benchmarking and evaluation without extra tooling.
• Creation/Creator — Custom preset authoring mode. Lets developers inspect the current execution environment, test Cordis plugins in memory, and combine them to build new tailored agent configurations.

All four modes run on the same plugin core — the difference is purely in which plugins are loaded at startup .

Deployment on China's National Supercomputing Internet

On August 14–15, 2026, China's National Supercomputing Internet made DeepSeek-V4-Pro-0813 and DeepSeek Harness v0.1 officially available on its platform — the country's first large-scale integrated computing resource pool with over 100,000 AI accelerator cards . This gives domestic developers access to model deployment, distributed inference, private hosting, and agent development within a single national computing infrastructure, lowering technical barriers for AI adoption in scientific research and industrial applications .

Competitive Positioning Against Claude Code

DeepSeek explicitly positions Harness as a rival to Anthropic's Claude Code and OpenAI Codex, targeting coding and office productivity use cases . The framework team lead, Cui Tianyi, opened global beta testing on August 2, 2026, and attracted dozens of developers before the public launch . The V4-Pro-0813 model also posted competitive agent-focused benchmark scores: Terminal Bench 2.1 at 87.9, DeepSWE at 62.7, Toolathlon-Verified at 74.1, and DSBench-FullStack at 71.1 .

API Price Increases — Reversal of Low-Cost Strategy

On the same day as the Harness launch, DeepSeek announced major API price increases that reverse its previous reputation as the "Token Price Butcher" :

• V4-Pro-0813 is priced at $1.32 per million input tokens and $3.96 per million output tokens — up to 14x higher than the V4 Flash model
• Increases range from 50% to 1,100% above prior prices, depending on the model, token type, and time of use
• A new peak/off-peak pricing structure was introduced, with peak rates rising as much as 12x on cache-hit input tokens — peak hours in China are 9:00–12:00 and 14:00–18:00 Beijing time, with off-peak rates set at half the peak rate
• DeepSeek had signaled a "significant" price hike on August 6, and analysts describe this as the end of the "bargain era" for the company's API services

Why This Matters

The simultaneous releases of Harness and V4-Pro-0813, alongside the pricing pivot, signal a strategic shift at DeepSeek: moving from a pure model provider to an execution-layer infrastructure company. By open-sourcing the agent runtime under MIT and making it available on national supercomputing infrastructure, DeepSeek is betting that the next competitive battleground in AI is not just the model, but the ecosystem of tools, agents, and workflows built around it.