Building AI agents remains challenging due to SDK abstraction limitations, especially around tool use and provider-specific features. Direct platform SDKs offer better control than higher-level abstractions. Explicit cache management (like Anthropic's approach) provides more predictability than automatic caching. Reinforcement
Table of contents
Which Agent SDK To Target?Caching LessonsReinforcement In The Agent LoopIsolate FailuresSub Agents / Sub InferenceThe Use Of An Output ToolModel ChoiceTesting and EvalsCoding Agent UpdatesQuick Stuff I Read And FoundSort: