The video compares two free coding agents, DeepSeek Harness (DSH) and Pi (formerly Py), which are notable for not originating from Anthropic or OpenAI. Both agents have quickly evolved, with Pi reaching version 1.0 within a day and DeepSeek Harness launching a desktop app. In a significant terminal task test, Pi solved 66 out of 89 tasks, while DeepSeek Harness solved 65, showing very close performance. Both agents use the same underlying model library when not running their own models, but they differ in their approach to the agent harness—the system that manages the model, tools, prompts, and context handling.
DeepSeek Harness offers a desktop app for Mac and Windows that bundles everything needed for immediate use, making it user-friendly for those who prefer a graphical interface. In contrast, Pi is designed as a minimal, terminal-based agent with a small core set of tools and relies on extensions for additional functionality, appealing to users comfortable with command-line environments. Both are MIT-licensed and free, but DeepSeek Harness emphasizes a plugin-heavy approach with 26 tools available by default, while Pi starts minimal and encourages adding features via extensions.
When comparing performance using the same model, both agents performed nearly identically in various tests, with Pi slightly edging out in some scenarios and DeepSeek Harness in others. DeepSeek Harness was generally faster in task completion, but Pi was more cost-effective, spending less per successful task due to its smaller prompt size. Pi also supports a broader range of model providers and local models out of the box, including advanced features like virtual models that route requests to different real models dynamically, whereas DeepSeek Harness primarily focuses on Deep Six’s own models and API.
In terms of features available before installing extras, DeepSeek Harness leads with built-in plan mode, to-do lists, web search, workflows, and sub-agents, plus a sandbox environment that restricts what the agent can do for safety. Pi provides fewer tools by default and no sandbox, recommending users run it in containers for security. Pi 1.0 introduced a new tool integration method called code mode, which improves efficiency by running tool calls in a sandboxed JavaScript environment, reducing prompt size and cost.
Regarding control and community involvement, Pi has a larger contributor base and actively merges external pull requests, making it more open and community-driven. DeepSeek Harness, while having more GitHub stars, does not accept outside contributions and is still in developer preview with frequent breaking changes. Privacy-wise, DeepSeek Harness sends session logs and usage analytics to Deep Six’s API by default, with limited user control, whereas Pi’s analytics are off by default and more privacy-conscious. The video concludes by recommending DeepSeek Harness for users wanting an out-of-the-box sandboxed app with Deep Six’s models, and Pi for terminal users who want flexibility with models and local setups, noting that the landscape is rapidly evolving.
