LangChain Ships Deep Agents v0.7 with 65% Fewer Base Input Tokens
Summary
- • LangChain's Deep Agents v0.7 cuts base input tokens 65% (~6k → ~2k) by removing built-in system prompts, trimming tool descriptions 43%, and making TodoListMiddleware opt-in
- • Performance validated across autonomous, conversational, and long-context benchmarks on four models: gpt-5.6-luna, gemini-3.6-flash, claude-sonnet-4-6, and claude-opus-4-8
- • Key finding mirrors Anthropic's simultaneous context engineering update: modern models no longer need verbose prompts; tool schemas teach usage better than few-shot examples
- • For gpt-5.6-luna: tokens down 34%, cost down 15%, reward up 4%; TodoListMiddleware retained as opt-in for long multi-step tasks and less capable models
Details
Deep Agents v0.7 Released
LangChain ships a leaner, more configurable agent harness reducing base input tokens by 65% while maintaining comparable performance across major frontier models.
65% Fewer Base Input Tokens per Turn
Base input tokens drop from ~6k to ~2k per default-agent turn via three changes: removed built-in system prompt, trimmed tool descriptions by 43%, and made TodoListMiddleware opt-in.
Tool Schemas Beat Few-Shot Examples
Key finding: good tool schemas teach model usage more effectively than few-shot examples, which can narrow exploration. Repetition across system prompt and tool descriptions adds no reinforcement value.
gpt-5.6-luna: 34% Fewer Tokens, 15% Lower Cost, +4% Reward
Most improvements held across all tested models; claude-sonnet-4-6 was the exception, showing a cost increase traced to two challenging autonomous tasks in LangSmith analysis.
3-Category Eval Suite Validates No Performance Drop
New eval suite covers autonomous tasks (coding, data analysis), conversational (multi-turn with simulated user), and long-context (retrieval and reasoning). v0.7 ran against v0.6.12 baseline across all four models.
TodoListMiddleware Now Opt-In
Evals showed slightly better rewards and lower cost with todos disabled by default. Still recommended for: long multi-step tasks, less capable models, and UI-facing use cases where plan visibility matters.
Anthropic Cut 80%+ of Claude Code's System Prompt Simultaneously
Anthropic published updated context engineering guidance showing 80%+ prompt cuts for Claude Opus 5 and Fable 5 with no measurable drop in coding evals — independently converging on the same principle.
Source: LangChain Blog (July 29, 2026). Eval methodology and full results linked in post.
What This Means
Deep Agents v0.7 reflects a broader industry shift: as large language models become more capable, the scaffolding around them can be dramatically simplified without hurting quality. The 65% token reduction is immediately valuable for teams running agents at scale — it delivers direct cost savings and in some cases improved performance with no application changes required. The simultaneous convergence with Anthropic's context engineering findings signals this is an emerging consensus rather than a single company's optimization, suggesting teams still using verbose prompting strategies with heavy system prompts and few-shot examples are leaving both performance and money on the table. For LangChain users specifically, upgrading to v0.7 is a straightforward win with asymmetric risk.
Sources
- Deep Agents v0.7Langchain