The DeepSeek Harness agent environment, built on the Cordis architecture, allows for autonomous project development through a plugin-based system. It enables agents to manage entire directories, execute commands, run tests, and utilize sub-agents for specific tasks like interface design or testing.
During an experiment to build a browser-based arcade game, the agent completed 183 steps over 2 hours, 43 minutes, and 27 seconds. The process consumed 35.7 million tokens, with 35.2 million attributed to cached input, significantly reducing costs compared to processing new data.
The project utilized the DeepSeek V4.1 Flash model via an OpenCode Go subscription, which costs $10 per month. The service offers varying rates for Peak and Off-Peak usage, with the latter providing lower costs for both input and output tokens.
The initial game, Type Strike, was developed as a local React application using TypeScript and Canvas. A subsequent task added server-side functionality, including PostgreSQL integration, Telegram authorization, and cloud-based statistics, resulting in the web application Typer Cool.
Despite the agent's capability to handle complex tasks, manual intervention was required for real-world deployment. Issues included mobile browser rendering inconsistencies, port conflicts in Docker environments, and strict API validation errors that were not caught during initial testing.
