DeepSeek launches Harness v0.1 open-source framework for AI agents

DeepSeek has launched DeepSeek Harness v0.1, an open-source, MIT-licensed agent framework designed to turn AI models into autonomous coding agents. Built on the Cordis meta-framework, Harness treats all components—including models, sandboxes, and filesystems—as swappable plugins. The launch coincides with the release of the agent-optimized DeepSeek-V4-Pro model and a major API pricing overhaul. Beginning August 16, 2026, DeepSeek will implement peak and off-peak rates, raising API costs by 50% to over 1,100% and shifting its competitive strategy from raw token pricing to agent orchestration infrastructure.

DeepSeek Harness v0.1 launch

• ▪DeepSeek Harness v0.1 is released under the permissive MIT license, with its source code made available for download on GitHub.
• ▪The DeepSeek Harness project is led by Cui Tianyi, who joined DeepSeek in March 2026 after previously working at the quantitative trading firm Jane Street.
• ▪DeepSeek launched a developer preview of DeepSeek Harness v0.1, an open-source agent runtime framework designed to turn AI models into autonomous coding agents, on August 13, 2026.

Plugin-based agent architecture

• ▪DeepSeek Harness supports chain of thought traces by recording system prompts, reasoning, tool calls, and context injections in an append-only session log visible in a Trajectory view.
• ▪DeepSeek Harness provides four operational settings, including a standard mode for general tasks, a code-focused mode, a creative mode, and a minimal mode for isolated testing.
• ▪DeepSeek Harness is built on Cordis, a meta-framework designed around composable plugins where models, tools, skills, sandboxes, and user interfaces are implemented as swappable components.
• ▪A research paper by Yifan Shi, Wei Zhang, and Tianyi Cui describes how the Cordis framework supports temporal and spatial composability to allow dynamic plugin modification without application crashes.

Competition with Anthropic Claude Code

• ▪DeepSeek Harness is positioned as an open-source, model-agnostic alternative to Anthropic's Claude Code and OpenAI's Codex agent infrastructures.
• ▪Unlike Anthropic and OpenAI, which restrict or summarize raw chain of thought reasoning, DeepSeek provides raw thinking mode by default on its API and displays it in DeepSeek Harness.

DeepSeek V4-Pro API pricing changes

• ▪Under the new pricing structure, DeepSeek-V4-Pro cache-miss input rises from $0.435 to $0.66 off-peak and $1.32 peak per million tokens, while output rises from $0.87 to $1.98 off-peak and $3.96 peak.
• ▪Beginning August 16, 2026, at 16:00 UTC, DeepSeek is replacing its flat API pricing with peak and off-peak rates, resulting in price increases ranging from 50% to over 1,100%.
• ▪DeepSeek launched the official version of its 1.6-trillion-parameter DeepSeek-V4-Pro model on August 13, 2026, featuring enhanced agent capabilities and native support for the OpenAI Responses API.

Developer adoption strategy

• ▪DeepSeek's transition to the MIT-licensed Harness framework represents a strategic pivot to establish developer lock-in at the agent orchestration layer as model intelligence becomes commoditized.
• ▪On its launch day of August 13, 2026, the DeepSeek Harness GitHub repository attracted significant developer attention, accumulating approximately 27,500 stars and 2,000 forks.