Agent Experience Stack (developer.microsoft.com)

A new article series on Agent Experience (AX) explores how AI coding agents can be made more effective within developers' tech stacks. While AI agents aim to boost productivity by generating code, they often miss the mark due to issues such as outdated SDKs or incorrect service selections. The series emphasizes understanding the layers between a developer’s prompt and the generated code, which include the model, the harness (the agent itself), and agent extensions. Developers have control primarily over agent extensions, which are crucial for guiding the model towards accurate outcomes. The significance of this series lies in its potential to enhance the interoperability of AI coding agents with diverse technologies by illuminating how agent extensions can correct biases inherent in the models and contend with the limitations of the harnesses. Key challenges discussed include discovery failure, where extensions are not recognized; selection failure, where the model misinterprets developer intent; and quality failure, where the output worsens the situation. By measuring the effectiveness of these extensions through controlled comparisons, developers can optimize their use of AI agents, targeting improvements in lift—where outcomes enhance productivity—while minimizing drag, which represents inefficiencies and potential disruptions in the coding process.