Deep Agents v0.7 cuts base tokens 65%

LangChain’s Deep Agents v0.7 cut the default harness from roughly 6,000 to 2,000 base input tokens, a 65% reduction, by removing the hidden system prompt, trimming tool descriptions 43%, and making todo middleware opt-in. LangChain’s four-model evaluation found overall reward held steady—with confidence intervals spanning zero—while GPT-5.6 Luna used 34% fewer tokens and cost 15% less. For teams competing in the quiet war over agent harnesses, the operator lesson is to benchmark inherited scaffolding before adding more: modern models may treat yesterday’s helpful prose as today’s context tax.