IBM experiments link context to coding assistant actions

The author explores the trade-offs of automatic model routing in coding assistants, arguing that while cheaper models can save API budget on routine steps, developers need visibility and control when investigations get stuck. Automatic routing can reduce costs on benchmarks, but hidden model selection deprives developers of valuable debugging context and the ability to escalate to a more capable model. The author also discusses ongoing experiments at IBM to connect context information with recorded actions to make AI assistant conversations easier to inspect. Open article →