This research proposes a novel approach to software agents by giving LLMs direct access to source code and runtime context, enabling them to inject and execute generated code for better performance.

• Addresses limitations of current API-based and GUI-based agent approaches
• Enables deeper software interaction through access to internal components
• Allows dynamic code injection, improving accuracy and efficiency
• Creates a blueprint for transforming traditional software into intelligent agents

This engineering breakthrough could fundamentally change how we design software assistants, creating more capable AI helpers that understand and manipulate programs at their core rather than just through external interfaces.

Original Paper: Every Software as an Agent: Blueprint and Case Study