Building Long-Running Agents with ScaleDown and Deep Agents

Building long-running agents comes down to the art of context management: deciding which information goes into the context window, and when. That means summarizing old history, compacting large tool results, and extracting the right data from long inputs. Each of those steps is a decision. Historically we’ve made them with LLMs, which is slow and expensive at the volume a long-running agent needs.

To make those decisions faster and cheaper, we now have a langchain-scaledown integration and support in the LangSmith LLM Gateway for ScaleDown’s constrained decoding models (CDMs). Where a general-purpose LLM writes any string one token at a time, a CDM is trained for one task and only returns answers that task allows, like a label from your list or spans of the input. That lets it be much smaller, faster, and cheaper: ScaleDown reports its models are 200x faster and 467x cheaper than frontier models on these narrow tasks.

The integration ships with three middlewares for LangChain and Deep Agents: one summarizes old history, one compresses large tool results, and one extracts structured data from the conversation. In this post, we cover what ScaleDown’s models are and how to use them to scale down your agent’s context.

What models does ScaleDown offer?

ScaleDown has four models, each trained for a single task, so each can be far more efficient than a general-purpose LLM:

• Summarize writes a short summary of long text.
• Compress cuts a block of context down to the parts relevant to a question.
• Extract pulls the fields you describe out of text, each with a confidence score and its position in the source.
• Classify scores text against labels you define and returns a probability for each. You can decide whether to return a natural language reason for why the classification was made.

Every model takes up to 1M tokens of input, so a full contract, a long policy page, or an hour of agent history fits in one call with no chunking. They’re also multimodal: Extract, Summarize, and Classify take images and scanned PDFs directly in addition to unstructured text, so a screenshot or a scanned invoice works as an input.

On ScaleDown's own benchmarks, measured 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, while scoring higher on accuracy for both.

If you’ve used a decision model like Jev, ScaleDown Classify will feel familiar: both return a typed answer with a probability. Jev takes text only, up to 32K tokens of state per question. Classify takes up to 1M tokens, images included, and adds a one-sentence reason to each decision. When an agent routes a ticket or stops a run based on a label, the reason tells you why, which makes the agent easier to debug.

Summarizing message history

Long-running agents need summarization because their history eventually outgrows the context window. A coding agent that has been working for an hour has dozens of file reads and tool outputs in its history, most of which it no longer needs verbatim. Summarizing them with another frontier model call adds cost and latency every time the history fills up.

ScaledownSummarizationMiddleware replaces Deep Agents’ built-in summarization middleware. It takes the same trigger and keep options to decide when to summarize and which recent messages to keep, and ScaleDown writes the summary:

The agent keeps its own model for planning and writing code. Only the summary moves to ScaleDown, which by default returns the same sections as LangChain’s summary prompt (intent, summary, artifacts, and next steps). If a ScaleDown call fails, the middleware leaves history as is and the agent keeps running. The same middleware works with LangChain’s create_agent.

Compressing tool results

Long-running agents call a lot of tools, and every large result stays in context for the rest of the run. A fetched web page can run to tens of thousands of characters when the agent only needs one paragraph of it. Every later step pays for those tokens again, and the agent gets worse at staying on task as its context fills up.

ScaledownToolCompressionMiddleware adds an optional compression_query parameter to your tools. When the model only needs part of a result, it fills in compression_query on the tool call, for example "refund policy for EU orders". The middleware strips it before the tool runs, then ScaleDown compresses the output to the relevant parts before the model sees it:

Your tools don’t change, so this works with any tool, including MCP tools. In this trace, ScaleDown cut a long docs page by almost 90% before the model saw it. The compressed result starts with a note like this one, so the model can call the tool again without compression_query if it needs the full page:

Extracting structured data

Agents constantly pull specific values out of what they read, like an ID, a date, or a field the next tool call needs. That extraction sits inside the agent loop, so it has to be fast enough not to add latency to every turn, and exact enough that a tool can use the value as is.

For example, before a support agent can file a ticket, it needs the order number, the error message, and the customer’s email. The main model could fill those in as tool arguments, but it may paraphrase an error code or mistype an email address instead of copying it exactly.

ScaledownExtractionMiddleware hands this step to ScaleDown. You describe each field in plain language, and Extract copies the value from the text. A field can also be a set of labels, which ScaleDown answers with Classify, so the ticket can be routed by issue type. Before the first model call of a run, the middleware sends the user’s messages to Extract and saves the fields to agent state under scaledown_extraction, where your tools can read them:

Extract only reads the user’s messages, never the agent’s replies, so every value is something the customer actually said. With a checkpointer, extraction re-runs each turn, so the record fills in as the customer adds detail. For "Order A-1042 failed with ERR_PAYMENT_DECLINED. I already tried a different card.", the fields come back as:

Classifying drift

An agent decides its next step based on its current state, like whether the task is done or whether the last tool call worked. Classifying that state happens on every turn, so it needs to be cheap and fast.

One state worth catching in long-running agents is drift: looping on the same failing tool call, or getting stuck in an unproductive cycle. There’s no prebuilt middleware for this, but you can call any ScaleDown model from your own custom middleware. Here’s a circuit breaker that runs after every model call, sends the history to ScaleDown Classify, and stops the run when it’s off track:

Classify returns a probability for each label, so you choose how confident it has to be before the breaker trips. Its one-line reason shows up in the final message, so you can see why the run stopped. A gentler version doesn’t end the run on the first trip: it adds the reason to the conversation as a nudge to change course, and only stops if the agent drifts again.

These four sections show a few places ScaleDown’s models fit, and the list isn’t exhaustive. You can mix and match them, like running extract / summarize on long tool results, or classification for runtime model routing. Check out ScaleDown Cookbooks to see other ways to use CDMs models across different types of data.

ScaleDown in the LangSmith LLM Gateway

The LangSmith LLM Gateway lets you call models across providers with one LangSmith API key, traces every call, and enforces your spend and rate limits. ScaleDown is now one of its direct model access providers, at the /scaledown path.

An admin adds your ScaleDown key to the workspace’s provider secrets as SCALEDOWN_API_KEY once. Then point langchain-scaledown at the gateway and authenticate with your LangSmith key:

Each ScaleDown call then shows up as a LangSmith trace and counts toward the same limits as your other model calls.

When not to use ScaleDown

Keep a general-purpose LLM for open-ended generation, like writing code, drafting replies, and planning. This follows the "cheap by default, frontier on exception" approach from our Jev posts: use a large general-purpose LLM for the hard parts, and specialized models for the high-volume steps around it.

ScaleDown is fast and cheap enough to run on every turn, so your agent can make the structured decisions that context management depends on without slowing down the main loop or paying frontier prices for them. In a long-running agent, those decisions happen on almost every turn, so the savings add up fast.

Get started

Install the integration and set your ScaleDown API key:

New accounts get 50M free tokens. The ScaleDown integration docs cover the configuration options for each middleware. Give langchain-scaledown a try, and let us know what you think on the LangChain forum, X, or GitHub.