On the evening of August 13th, DeepSeek made two significant announcements: it officially released DeepSeek V4 Pro, accessible via an application, web interface, and API, and introduced the first post from the DeepSeek Harness team, featuring a 'black whale' avatar, announcing the preview version of Harness, which was released as open-source under the MIT license.
The significance of Harness lies in the fact that in the era of agents, the model is no longer the sole factor determining an AI's ability to perform tasks. Previously, on August 6th and 11th, Composio, a company specializing in agent tools, integrated the same DeepSeek V4-Flash model into eight different execution systems (harnesses). These systems were requested to complete 30 multi-step tasks in real applications such as Gmail, Google Calendar, GitHub, and Slack.
The results of these tests showed significant discrepancies: the best system completed 20 tasks, while the worst managed only 14. Out of 240 total runs, only 129 were successful, and only six out of 30 tasks were executed by all eight systems. In terms of cost, running 16 tasks using Claude Code, Codex, and DeepAgents cost approximately $0.195, $0.081, and $0.045 per successful task, respectively, demonstrating a fourfold difference in cost when using the same model.
According to DeepSeek's internal stance, agents must solve the problem of continuous learning, which will ultimately allow AI to accelerate its own development. Harness is positioned as the working environment through which models enter real development. The company, founded on low cost and model strength, cannot allow task execution and developer entry points to depend on external systems.
The core design principle, according to DeepSeek's documentation, is that everything is a plugin. Harness is built on the Cordis plugin system: models, tools, skills, sessions, sandboxes, storage, cycles, planning, and user interface—all are formed from plugins. This differs from standard plugins: VS Code plugins add features to the editor, Codex plugins add tools to the agent, whereas Harness plugins can completely replace the agent's brain, toolset, rules, or appearance.
During an internal review, developers were able to create hundreds of plugins in just a few days, including one that implemented long-term memory between sessions with background self-evolution. Another key feature is the ability to track history. Harness uses session logs with append-only session logs: everything the model sees—system prompts, reasoning chains, tool calls, and results, sub-agent planning, and every context injection—is fully recorded, and context compression never deletes the original history. Developers can view the history by source in the Trajectory view with recovery, forking, extraction, and replay capabilities.
The team lead, Qiu Tianyi, holds a computer science degree from Zhejiang University and worked for nine years at the quantitative trading firm Jane Street before joining DeepSeek in March 2026, when the Harness project internally began. Harness natively supports nearly 40 major models, including Kimi, OpenAI, Anthropic, and Google, positioning itself as a set of components rather than a finished agent. The v0.1 version number clearly indicates that the 'black whale' has only just surfaced.
