The framework lets models work directly with code repositories: locating files, modifying code, running tests and iteratively fixing errors.
DeepSeek has released Harness, its first agent runtime framework. It gives models direct access to code repositories: locating files, changing code, running tests and correcting errors across several passes.
That moves the point of application. Until now models produced suggestions that a person adopted. A runtime executes the steps itself and checks the result against tests. Its usefulness therefore depends on the quality of existing test coverage.
For companies with in-house software development this is the practically relevant question. Without dependable automated tests such an agent lacks a corrective and changes surface late. With good coverage, the effort for routine work falls markedly.
What this means for decision-makers
- Judge the value of code agents by your test coverage, not by demonstrations.
- Define which repositories an agent may write to and keep them separate from production branches.
- Require a traceable change history per agent run before approving such tools.
This story was produced automatically from the source named above and checked by software before publication. The image is symbolic and shows neither the event nor a real person. How this paper is made