DeepSeek Expands AI Agent Foundation

DeepSeek, a Chinese AI research organization, has simultaneously released its new flagship model 'DeepSeek-V4-Pro' and an open-source agent execution framework 'DeepSeek Harness (dsh) v0.1'. V4-Pro is an updated version emphasizing agent workloads and is available through three channels: web interface, mobile app, and API. It also supports integration with OpenAI's Responses API and Codex.

DeepSeek has demonstrated its presence by providing high-performance models at low prices. However, this move goes beyond merely offering models and enters the realm of building 'systems' that enable developers to actually operate AI. An AI agent refers to a system of AI that follows instructions, uses multiple tools, and autonomously performs tasks such as file manipulation and web search. The execution framework responsible for designing and managing such agents corresponds to the Harness released this time.

The design philosophy of Harness is distilled into a single point: 'make everything a plugin.' Models, tools, session management, sandboxes, file systems, loops, orchestration (coordination of multiple processes), and UI are all structured so that developers can freely replace and extend them. It is built on Cordis, a plugin-based framework, and is available under the MIT license from GitHub. It can be launched with the npm command 'npx @deepseek-ai/dsh web'. However, it is in a developer preview stage, and the official repository explicitly states that 'changes that break compatibility may occur.'

Significant changes are being introduced in pricing. DeepSeek is switching its API pricing structure from a uniform price to a time-based 'peak and off-peak system'. This change becomes effective on August 17, 2025, at 1:00 AM Japan time (UTC 16:00). Even discounted prices during off-peak hours are expected to significantly exceed current price levels, meaning V4-Pro API usage costs will increase overall.

DeepSeek's competitiveness has previously lay in a combination of high performance and low pricing. This price increase can be seen as a departure from that strategy, but it can also be viewed as adding a new differentiator with Harness. Anthropic's Claude Code and OpenAI's Codex are gaining adoption among developers as coding agents, but they are provided as proprietary, closed environments. In contrast, Harness employs a different strategy, releasing an open-source layer that provides equivalent functionality.

What DeepSeek is aiming for can be understood as establishing presence not only in AI model competition itself, but in the 'overall ecosystem' including tools and execution environments surrounding the model. Harness is currently at the developer preview stage and is not positioned as an enterprise-grade stable release. However, if it becomes widely adopted as a modular open-source platform, the very basis on which developers make decisions about which models to use could change. The maturity and adoption status of Harness going forward will be a key point to watch.