Analysis Finds Claude Code Resends Full Background Agent Transcripts to Model

A DEV Community article examines how Claude Code's autonomous background agents maintain state across tasks. It reports that the system serializes the agent's full transcript and sends it to the model on each interaction to preserve context. The author argues this design treats the LLM as stateless and causes context inflation, raising token costs and latency as tasks grow. The piece also highlights potential security concerns about transmitting complete workspace histories to model providers. It suggests external state management, such as persistent stores or diff-based summaries, as a more efficient alternative.