DeepSeek has released a preview version of DeepSeek Harness, an open source tool for developing and running AI agents. The project relies on flexibility, allowing developers to independently change various components of the system.
In DeepSeek Harness, plug-ins are models, tools, skills, sessions, sandboxes, file systems, agent cycles, and interfaces. The core is the Cordis meta-framework, which supports dynamic addition and deletion of components without restarting the application. The system can also take into account dependencies between plugins, which should simplify the creation of complex configurations.
Of particular interest is the work of the recording agents. DeepSeek Harness stores system commands, tool calls, results, subagent work, and context changes in a single thread. The resulting recording can be viewed, searched, played, and used to continue or branch the session. This approach provides developers with more opportunities to analyze the behavior of the model.
The release of this tool shows that the competition in artificial intelligence is increasingly turning to more than just the models themselves. Agent shells determine how to use them, and familiar settings and tools can tie developers to specific systems. Pi creator Armin Ronacher has cited DeepSeek Harness as an inspiration and admitted that he wanted to reconsider some of his decisions after becoming familiar with the project.
Please write in the comments which AI model you prefer.