---
type: "post"
id: "building-long-running-agents-with-scaledown-and-deep-dd78b3"
slug: "building-long-running-agents-with-scaledown-and-deep-dd78b3"
title: "Building Long-Running Agents with ScaleDown and Deep Agents"
description: "The thesis is: Selected context-management work is being split out from the general-purpose model that runs an agent."
retrieval_nugget: "The thesis is: Selected context-management work is being split out from the general-purpose model that runs an agent."
published_at: "2026-10-13"
updated_at: "2026-10-13"
record_date: "2026-10-13"
date_kind: "published_at"
topics: []
source_urls: ["https://www.langchain.com/blog/building-long-running-agents-with-scaledown-and-deep-agents"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:building-long-running-agents-with-scaledown-and-deep-dd78b3"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:0f4b4cb8ed43d2efb1db2ee5237c66f5ec2e5f18bac6605517dd77d28894924b","reviewed_at":"2026-10-13T12:04:03.189Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":2,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Selected context-management work is being split out from the general-purpose model that runs an agent.","observed_facts":[{"text":"The announcement describes middleware for summarization, tool-result compression, and structured extraction in LangChain and Deep Agents.","source_urls":["https://www.langchain.com/blog/building-long-running-agents-with-scaledown-and-deep-agents"]},{"text":"The announcement says ScaleDown is supported in the LangSmith LLM Gateway, with calls traced and counted toward spend and rate limits.","source_urls":["https://www.langchain.com/blog/building-long-running-agents-with-scaledown-and-deep-agents"]},{"text":"LangChain shows a trace reducing a documentation result from 18,402 to 3,110 characters.","source_urls":["https://www.langchain.com/blog/building-long-running-agents-with-scaledown-and-deep-agents"]}],"mechanism":"Narrow, constrained-decoding models handle bounded context operations while the agent’s main model continues to plan and execute the task.","why_now":"Long-running agents repeatedly accumulate history and tool output, so context handling is part of their ongoing execution cost and latency, not merely an initial prompt-design choice.","implications":["Teams can test context operations as replaceable middleware without changing the model responsible for the agent loop.","A trace-level reduction in tool output is evidence of the described flow, not proof of improved end-to-end agent performance."],"evidence_boundary":"This is a LangChain product announcement, and its comparative speed and cost claims are attributed to ScaleDown rather than independently verified.","watch_conditions":["Observe whether teams adopt these middleware in production and whether measured end-to-end quality, latency, and cost improve across real workloads."],"related_records":[],"new_branch_reason":"This starts a distinct analytical record because no prior New Runtime record was supplied for this specific integration."}
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/building-long-running-agents-with-scaledown-and-deep-dd78b3/","markdown":"https://newruntime.com/posts/building-long-running-agents-with-scaledown-and-deep-dd78b3.md","json":"https://newruntime.com/posts/building-long-running-agents-with-scaledown-and-deep-dd78b3.json"}
---

# Building Long-Running Agents with ScaleDown and Deep Agents

## Retrieval answer

The thesis is: Selected context-management work is being split out from the general-purpose model that runs an agent.

The thesis is: Selected context-management work is being split out from the general-purpose model that runs an agent.

LangChain’s announcement describes a `langchain-scaledown` integration for LangChain and Deep Agents, with middleware for summarization, tool-result compression, and structured extraction. It also says ScaleDown calls are supported in the LangSmith LLM Gateway and appear in tracing and spend and rate-limit accounting. The mechanism is: Narrow, constrained-decoding models handle bounded context operations while the agent’s main model continues to plan and execute the task.

The announcement says the integration can summarize older history, compress tool outputs before the agent sees them, and extract fields from user messages into agent state. LangChain gives one trace in which a documentation result shrank from 18,402 to 3,110 characters. That example demonstrates the described flow, not a general performance guarantee. Why now: Long-running agents repeatedly accumulate history and tool output, so context handling is part of their ongoing execution cost and latency, not merely an initial prompt-design choice. The first implication is: Teams can test context operations as replaceable middleware without changing the model responsible for the agent loop.

The evidence boundary is: This is a LangChain product announcement, and its comparative speed and cost claims are attributed to ScaleDown rather than independently verified. The first watch condition is: Observe whether teams adopt these middleware in production and whether measured end-to-end quality, latency, and cost improve across real workloads. This starts a distinct analytical record because no prior New Runtime record was supplied for this specific integration.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)