Long-running agents stop paying frontier prices for context chores: ScaleDown models land in LangChain
Synopsis
Long-running agents previously made every context-management decision with a general-purpose LLM, which is slow and expensive at the volume such agents need. ScaleDown's constrained decoding models (CDMs) are trained for one task and only return answers that task allows, so they can be smaller, faster and cheaper; a new LangChain and Deep Agents integration ships three middlewares for summarization, tool-result compression and structured extraction, and the LangSmith LLM Gateway now lists ScaleDown as a direct model provider so calls are traced and count toward spend and rate limits. ScaleDown reports its models are 200x faster and 467x cheaper than frontier models on these narrow tasks.
Interpretation
Every context-management decision a long-running agent makes each turn can now go to a task-specific model instead of a general-purpose LLM. Summarizing old history, compacting large tool results and extracting fields from long inputs were previously done by general-purpose LLMs, which is slow and expensive at the volume a long-running agent needs. ScaleDown reports its constrained decoding models are 200x faster and 467x cheaper than frontier models on these narrow tasks; on ScaleDown's own benchmarks against the average of GPT-5.5, GPT-5.4 Mini and GPT-5.4 Nano on public datasets, Summarize is 41x cheaper and 2.7x faster and Compress is 138x cheaper and 1.8x faster, both scoring higher on accuracy.
The integration splits context management into three pluggable middlewares for summarization, compression and extraction, while the agent keeps its own model for planning and writing code. A CDM differs from a general-purpose LLM: the general model writes any string one token at a time, while a CDM only returns answers its task allows, such as a label from a list or spans of the input. Four models each train on a single task: Summarize writes short summaries, Compress cuts context to the parts relevant to a question, Extract pulls described fields with a confidence score and source position, and Classify scores text against user-defined labels and returns a probability; each takes up to 1M tokens of input, and Extract, Summarize and Classify also accept images and scanned PDFs.
Perspective
The integration targets developers building long-running agents with LangChain and Deep Agents, covering four narrow tasks: summarization, tool-result compression, structured extraction and state classification, and it can be called through the LangSmith LLM Gateway at the /scaledown path so calls are traced and count toward spend and rate limits. Open-ended generation such as writing code, drafting replies and planning should stay with a general-purpose LLM.
The speed and cost multiples are reported by ScaleDown, and its own benchmarks compare against the average of GPT-5.5, GPT-5.4 Mini and GPT-5.4 Nano; the material gives no third-party replication. The compression example cuts one docs page from 18,402 to 3,110 characters, an observation from a single trace. The drift circuit breaker requires writing custom middleware, and no prebuilt implementation is provided.