DeepSeek’s Double Launch: V4‑Pro Model Goes Live and Harness Open‑Source, Advancing Agents

On August 13, DeepSeek delivered two announcements in one day: the silent rollout of the flagship DeepSeek‑V4‑Pro model and the open‑source release of the DeepSeek Harness runtime, together forming the story of "Agent = Model + Harness".

The official website highlights the formula Agent = Model + Harness , a concept originally seen in a LangChain blog post titled "The Anatomy of an Agent Harness".

Model Half: V4‑Pro Formal Release and Agent Leap

The API documentation quietly changed to version DeepSeek‑V4‑Pro‑0813, and users discovered the update on social media. The V4‑Pro model enhances agent capabilities, especially in production environments.

Pricing: Output costs 6 CNY per million tokens, input 3 CNY (cache hit as low as 0.025 CNY), roughly 1/60 of competing services. Starting August 17, peak‑off‑peak pricing applies: peak input 9 CNY, output 27 CNY per million tokens, with off‑peak rates halved. The cache‑hit price increase amounts to a 12‑fold jump.

Key specifications include 1 M token context, up to 384 K output, three reasoning modes (low / high / max), expert mode for app/web calls, and native OpenAI‑compatible API responses with Codex‑specific adaptations.

Some users report that the new model’s reasoning chains produce fragmented grammar and omitted words, likely due to token compression trade‑offs that favor agent inference over writing quality. Independent benchmarks (e.g., Artificial Analysis score 45.3) differ from the vendor’s claims, indicating the need for further verification.

Harness Half: DeepSeek Harness Open‑Source, "Everything Is a Plugin"

On the evening of the V4‑Pro launch, DeepSeek released the developer preview DSH v0.1 under the MIT license on GitHub, quickly gathering over 64 k stars (as of August 14).

Harness is described as a runtime that turns a model into a "working agent". Its design philosophy is "Everything Is a Plugin"—model, tools, skills, conversation, sandbox, storage, agent loop, scheduler, UI, etc., are all packaged as interchangeable plugins.

The only non‑plugin component is the Cordis kernel, which manages plugin loading, unloading, and dependency handling. Cordis provides two composability guarantees:

Temporal composability: ensures that a plugin’s side effects are fully reversible upon unloading.

Spatial composability: allows a plugin that depends on others to dynamically adapt when those dependencies appear, disappear, or change.

These guarantees enable true hot‑plugging: plugins can be added or removed during an agent’s execution without losing state, supporting a vision of agents that continuously install new capabilities and evolve autonomously.

The Cordis design is documented in DeepSeek’s accompanying paper.

Getting Started (30‑Second Guide)

Install Node.js, then run a single command: npx @deepseek-ai/dsh web On first run, provide your DeepSeek API key; the browser will open http://127.0.0.1:3080.

Four preset modes are offered, each representing a different plugin composition template:

Standard Mode: full toolset with code‑agent capabilities (file editing, shell, search, skills, sub‑agents, workflows). Recommended for first‑time users.

PTC Mode (Programmatic Tool Calling): the model generates TypeScript code that batches multi‑turn tool calls into a single execution, reducing token usage. Suited for structured, parallelizable tasks but requires strong model coding ability.

Minimal Mode: only a shell and a file editor, used for baseline model benchmarking; DeepSeek used this mode to evaluate V4‑Pro.

Creative Mode: lets the agent inspect its Cordis environment, experiment with plugins in memory, and dynamically create new plugins on the fly—analogous to an agent fabricating a missing wrench during work.

Each run generates an append‑only event log capturing system prompts, reasoning chains, tool calls, sub‑agent scheduling, and context injections. The Trajectory view lets developers trace where an agent deviated, providing observability, auditability, and reproducibility.

Harness does not lock developers into DeepSeek models; it supports any model listed in the official catalog, custom base URLs, and protocols, enabling integration with GLM, Anthropic, OpenAI, and others.

Implications

Model Layer: The V4 series reaches SOTA performance, bringing open‑source models to parity or beyond top‑tier closed models in agent tasks.

Execution Layer: Open‑sourcing Harness shifts DeepSeek from merely selling models and APIs to providing the execution environment that powers Claude Code and Codex, completing the Vibe Coding stack.

Ecosystem Layer: The "Everything Is a Plugin" MIT‑licensed approach invites global developers to co‑build the runtime; peak‑off‑peak pricing balances compute costs with commercial returns.

Critics note a steep learning curve, limited out‑of‑the‑box functionality, and a rough user experience, describing the product as a "diode"—appealing to enthusiasts but inaccessible to casual users. DeepSeek appears unconcerned, echoing its slogan "Exploring Uncharted Territories".

On the same day, competitors such as xAI released Grok 4.6 and Alibaba open‑sourced Qwen 3.8‑2.4T‑A95B, signaling a shift in large‑model competition from pure chat ability to practical agent execution.

In DeepSeek’s view, "Model sets the ceiling, Harness sets the floor"—controlling both ends of the agent stack.