DeepSeek has made its biggest move yet toward autonomous AI agents, pairing a new open-source framework with the official release of its updated V4-Pro model. The result is a platform aimed at developers who want to build agents from interchangeable parts instead of locking the entire runtime into one fixed design.
On August 13, 2026, DeepSeek released DeepSeek Harness v0.1 and DeepSeek-V4-Pro. Harness enters developer preview under the MIT license, while V4-Pro moves beyond its earlier preview and becomes available across the web, mobile app, and API.
A Modular Foundation for AI Agents
DeepSeek Harness is built around a striking idea: practically every part of an agent runtime can be swapped out as a plugin. That design gives developers room to change how an agent handles tools, workflows, and other runtime functions without rebuilding the entire system from scratch.
DeepSeek describes Harness as an open-source agent harness built on Cordis, a framework designed around composable plugins. Instead of treating an agent as one sealed product, the framework breaks its operation into pieces that developers can replace and combine.
The code is available now for download and use on GitHub, where DeepSeek’s repository was attracting roughly 27,500 stars and 2,000 forks as of August 13. That early activity gives the developer preview a strong launch signal, especially as open-source teams look for more control over agent infrastructure.
There is an important warning attached to the release: Harness remains a developer preview and will include compatibility-breaking changes. Developers can start experimenting now, but the framework’s interfaces may shift as DeepSeek builds out the platform.
V4-Pro Makes the Agent Push Concrete
DeepSeek introduced the V4 family in preview in April, with two models designed for different levels of scale. The lineup includes the 1.6-trillion-parameter V4-Pro and the 284-billion-parameter V4-Flash, and both models support context windows of up to one million tokens.
V4-Pro has 49 billion parameters activated per token, while V4-Flash has 13 billion parameters activated. The August 13 release marks the transition from the earlier preview into an updated official version of V4-Pro, with DeepSeek placing agent performance at the center of the launch.
DeepSeek reports that V4-Pro-0813 scored 87.9 on Terminal Bench 2.1, 74.1 on Toolathlon-Verified, 71.1 on DSBench-FullStack, and 67.2 on DSBench-Hard. Those benchmarks point directly at tasks involving tools, terminals, and software development, matching the model’s role inside an agent-focused platform.
V4-Pro also adds native OpenAI Responses API support. DeepSeek says the model is optimized for OpenAI’s own open-source harness, Codex, with one-click setup, giving developers a direct path from existing agent workflows to the updated model.
More Control Over Reasoning and Cost
DeepSeek is making reasoning effort an explicit control across both V4-Pro and V4-Flash. Developers can choose between three levels: Non-think, Think High, and Think Max.
That control lets users decide how much reasoning effort a task should receive instead of applying one setting to every request. For agent systems, where simple actions and complex multi-step work can share the same model, the three-level approach creates a clear way to tune behavior.
The pricing also separates the two models. V4-Flash costs $0.14 per million cache-miss input tokens and $0.28 per million output tokens. V4-Pro costs $0.435 per million cache-miss input tokens and $0.87 per million output tokens.
DeepSeek will begin implementing peak and off-peak API rates on August 16, 2026, at 16:00 UTC. That change adds another pricing factor for developers planning workloads around the API, especially when an agent may generate repeated requests across long-running tasks.
DeepSeek’s launch connects three pieces that are usually developed separately: a large language model, an agent runtime, and controls for reasoning effort and API use. Harness supplies the modular framework, V4-Pro supplies the updated agent-focused model, and V4-Flash offers a lower-priced option for developers balancing capability with token costs.
The developer preview will now show whether that modular design can support real projects as compatibility-breaking changes arrive. With the code available under the MIT license, the next stage belongs to developers who will test the plugins, shape the framework, and push DeepSeek’s V4 models into more autonomous systems.
Based on
- An open source rival to Claude Managed Agents just launched — thenewstack.io
- DeepSeek Harness launches as open source rival to Claude Code, alongside V4-Pro on API with higher prices | VentureBeat — venturebeat.com
- Launch of DeepSeek’s Harness marks its strategic pivot towards autonomous agentic AI | South China Morning Post — scmp.com
