DeepSeek Opens Its Agent Runtime With a Plugin for Everything

DeepSeek is opening up more than another model release. On August 17, 2026, the Chinese artificial intelligence company released DeepSeek Harness v0.1 in developer preview, publishing the full source code under the MIT license. The project ships as dsh at deepseek-ai/deepseek-harness.

The idea behind Harness is simple but ambitious: an AI agent needs more than a model. It also needs the tools, files, sandboxes, and control loop that let it keep working inside an environment. DeepSeek describes that relationship as Agent = Model + Harness, placing the runtime around the model at the center of the project.

DeepSeek announced Harness v0.1 alongside the official release of DeepSeek-V4-Pro on August 13. The developer preview followed on August 17, giving developers access to the open-source framework and its source code under the MIT license.

A different approach to building AI agents

Most agent harnesses hard-code the main parts of the runtime. The agent loop, tool registry, and session store remain fixed, while developers can extend the system only through the hooks its authors provide. Harness takes the opposite position, stating the project’s central idea in the first lines of its README: “Everything is a plugin.”

That rule reaches across nearly every part of the system. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the user interface all sit behind Cordis plugin boundaries. Developers can select, replace, or extend those parts through configuration without changing Harness source code.

DeepSeek describes Harness, or dsh, as an open-source agent harness built on Cordis, a meta-framework designed around composable plugins. That structure makes the model just one replaceable part of the runtime rather than the unavoidable center of a vertically integrated stack.

Standard mode shows what the framework is meant to handle. DeepSeek explicitly describes it as a full coding agent with file editing, shell access, search, planning, subagents, and workflows. Those capabilities sit alongside the project’s plugin-based approach, allowing the surrounding runtime to be replaced or extended as configuration changes.

The repository was already attracting developer attention on launch day. As of August 13, it had roughly 27,500 GitHub stars and 2,000 forks.

Harness arrives with DeepSeek-V4-Pro

The framework launch is tied to DeepSeek’s flagship model for agentic workloads. DeepSeek-V4-Pro has 1.6 trillion parameters, with 49 billion parameters activated per token, and supports context windows of up to one million tokens. DeepSeek says the model is now fully available across its web service, mobile app, and API.

DeepSeek’s changelog says the general-availability model has “significantly enhanced agent capabilities,” with a particular focus on production environments. The company also says V4-Pro is optimized for OpenAI’s open-source harness, Codex, with one-click setup.

DeepSeek published scores for V4-Pro-0813 across four tests: 87.9 on Terminal Bench 2.1, 74.1 on Toolathlon-Verified, 71.1 on DSBench-FullStack, and 67.2 on DSBench-Hard. These figures place the model’s agent-focused performance alongside the release of a runtime built to let developers change how an agent operates.

DeepSeek also lists DeepSeek-V4-Flash, which has 284 billion parameters with 13 billion activated per token. Both V4-Pro and V4-Flash support context windows of up to one million tokens.

Pricing gives developers another choice

DeepSeek announced API pricing changes beginning August 16, 2026, at 16:00 UTC. V4-Flash costs $0.14 per million cache-miss input tokens and $0.28 per million output tokens.

V4-Pro costs $0.435 for cache-miss input tokens and $0.87 for output tokens. The two pricing levels give developers a choice between the flagship model and the lower-priced V4-Flash while both support the same maximum context window.

DeepSeek’s release combines three pieces: an open-source runtime, a model aimed at agentic workloads, and API access with separate pricing for V4-Flash and V4-Pro. Harness makes the runtime configurable at nearly every layer, while V4-Pro supplies the flagship model that DeepSeek says has stronger agent capabilities.

For developers, the important shift is not only access to another coding agent. It is the chance to treat the model, tools, sessions, storage, scheduling, and interface as replaceable parts of one system. DeepSeek Harness v0.1 is still a developer preview, but its MIT license and plugin-first design make the project a clear move toward configurable autonomous AI agents.

Stay connected via Google News

Follow us for the latest AI News