LangChain ships Deep Agents v0.7, achieving a 65% reduction in base input tokens at comparable performance. Key changes include removing the hidden base system prompt, trimming built-in tool descriptions by 43%, and making TodoListMiddleware opt-in after evals showed it didn't meaningfully improve results. The release also adds full configurability over the middleware stack — users can now override defaults like SummarizationMiddleware with custom thresholds and prompts. Filesystem tools received optimizations including overwrite-on-write behavior, paginated read reporting, and truncation-safe grep/glob. Validated across four models (GPT, Gemini, Claude Sonnet, Claude Opus), the new harness generally reduces tokens and cost while holding reward steady.
