LangChain Deep Agents v0.7: 65% base-token cut via trimmed prompt harness; opt-in todos, middleware override

LangChain released Deep Agents v0.7, refactoring its base agent harness to cut input tokens by 65% (6k → 2k tokens per turn) while maintaining or improving task performance. Changes include: (1) removal of the default system prompt and general tool-guidance prose, (2) 43% trimming of built-in tool descriptions, (3) TodoListMiddleware now opt-in rather than default. The premise mirrors Anthropic's recent finding that Claude Code's system prompt was reduced >80% for Opus 5 / Fable 5 models without eval degradation, validating the thesis that modern frontier models are over-scaffolded.

Validation ran against a new eval suite spanning autonomous (coding, data analysis), conversational (multi-turn user interaction), and long-context (retrieval + reasoning) tasks across four models: GPT-5.6-Luna, Gemini-3.6-Flash, Claude Sonnet 4.6, and Claude Opus 4.8. GPT-5.6-Luna showed 34% token reduction and 15% cost reduction with +4% reward lift. Most models held steady on reward while dropping tokens/cost. Claude Sonnet 4.6 was an exception (cost up on two challenging autonomous tasks), flagged in LangSmith traces.

For platform teams: this is a production ops win. Reducing base harness tokens from 6k to 2k per turn means cheaper inference, lower latency, and more room for context on constrained deployments. The opt-in middleware approach (TodoListMiddleware, SummarizationMiddleware override) gives builders fine-grained control over prompt strategy without fighting the framework. The timing aligns with a broader pattern: as model capability stabilizes, harness-level efficiency becomes the lever.