LangChain Launches ScaleDown Integration to Enable Long-Running Agents
Current overview (not historical)
LangChain has released the ScaleDown integration, introducing four specialized models—Summarize, Compress, Extract, and Classify—to support the construction of long‑running agents. The new integration is designed to work as middleware within LangChain’s Deep Agents framework, allowing developers to offload heavy processing tasks and maintain agent responsiveness over extended periods. Following the ScaleDown announcement, a separate blog post addressed common questions about Agent Evals. It highlighted how to evaluate agents using trace analysis and small‑scale datasets, but did not directly tie these evaluation methods to the ScaleDown functionality. The two posts together outline both the technical enhancements for sustained agent operation and the practical evaluation strategies for developers working with LangChain agents.
Written with AI from the reporting · updated 10/15/2026
Related events
Expand the detailed fact route and sources
Story route
A station-by-station route from the first public evidence onward; every station carries its own evidence and uncertainty.
- 1Unclassified fact stationOct 13, 00:00recorded Oct 15, 02:23single sourcerevised
Building Long-Running Agents with ScaleDown and Deep Agents
Read original source ↗ LangChain BlogReplay this moment in the explorerEvidence & revisions · 1 source records
- Subject：
- LangChain
- Action：
- 发布 ScaleDown 集成和 LangSmith LLM Gateway 支持
- Object：
- ScaleDown 受限解码模型 (CDMs) 与 LangChain 深度代理
- Occurred (fact side)：
- Oct 13, 00:00
- Published (valid side)：
- Oct 13, 22:01
- First seen (known side)：
- Oct 15, 02:23
Fact stationSingle source · one-sided claimRevisedCopy may have been revised later; time replay is bounded. Layer labels group reading only, not causation between layers.
- 2Unclassified fact stationOct 15, 00:00recorded Oct 15, 09:26single sourcerevised
The Most Common Questions About Agent Evals, Answered
Read original source ↗ LangChain BlogReplay this moment in the explorerEvidence & revisions · 1 source records
- Subject：
- LangChain
- Action：
- 发布博客文章
- Object：
- 提供关于 Agent Evals 的常见问题解答
- Occurred (fact side)：
- Oct 15, 00:00
- Published (valid side)：
- Oct 15, 09:19
- First seen (known side)：
- Oct 15, 09:26
Fact stationSingle source · one-sided claimRevisedCopy may have been revised later; time replay is bounded. Layer labels group reading only, not causation between layers.
Uncertainty marks flag where to check the sources; they are not proven causation.
Report timeline
Explore every side through dated source reports.
- LangChain BlogRelated fact: single sourceSource report revisedLangChain Blog Answers Common Questions on Agent Evals for Agent Development
Report first recorded Oct 15, 09:26
Agent evals are necessary to check if an agent behaves as expected, as agents differ from traditional software because outputs vary between runs, multiple answers can be correct, and agents choose their own steps. LangChain's platform, LangSmith, guides users to start by understanding agent traces, building a small dataset (10 to 20 examples) from traces, running a baseline experiment, and establishing an improvement loop by adding new failures found during testing or production. Evaluation types include Code, LLM-as-a-judge, and Human Review.
Read original source ↗
- LangChain BlogRelated fact: single sourceSource report revisedLangChain integrates ScaleDown models to build efficient long-running agents
Report first recorded Oct 15, 02:23
ScaleDown offers four specialized models—Summarize, Compress, Extract, and Classify—designed for specific tasks, making them faster and cheaper than general-purpose LLMs. These models are integrated into LangChain and LangSmith LLM Gateway via middlewares to manage agent context by summarizing history, compressing tool results, and extracting structured data. ScaleDown models are reported to be significantly faster and cheaper than frontier models on narrow tasks, with Summarize being 41x cheaper and 2.7x faster than the average of GPT-5.5, GPT-5.4 Mini, and GPT-5.4 Nano on public datasets. All models support up to 1M tokens of input and are multimodal, accepting images and scanned PDFs.
Read original source ↗
Event discussion
Reader opinions are hosted on GitHub Discussions; posting requires a GitHub account. Comments are not verified facts.