Using AI agents to modify third-party libraries can be faster than submitting issues or pull requests. Developers can implement necessary features locally, avoiding the need to wait for maintainers to approve or release changes. However, maintaining a fork requires ongoing effort to keep it updated with the original repository.
Cohere has automated the maintenance of its vLLM fork using a set of instructions called vllm-skills. This process allows agents to perform auto-rebasing, resolve conflicts, and run tests. While this reduces the time required for updates from weeks to days, it still requires human oversight to define test parameters and verify the final results.
Forking does not guarantee that code will remain stable, as private versions are not tested by the broader community. For instance, an agent-generated fix for terminal table alignment in pyinfra initially failed to account for complex emoji, causing regressions that were only caught through manual review before the fix was merged into the main project.
The decision to maintain a fork depends on the cost of waiting for upstream changes versus the cost of internal maintenance. If a feature is specific to a single product, a fork may be justified. Conversely, if a fix is broadly useful, contributing it back to the original project remains beneficial for long-term stability and community support.
