自律的コンテキスト圧縮

imageTL;DR: We've added a tool to the Deep Agents SDK (Python) and CLI that allows models to compress their own context windows at opportune times.

Motivation

Context compression is an action that reduces the information in an agent’s working memory. Older messages are replaced by a summary or condensed representation of an agent’s progress that preserves what’s relevant to a task. This action is often necessary to accommodate finite context windows and reduce context rot.

Agent harnesses often control this by compacting at a fixed token threshold (deepagents uses model profiles to compact at 85% of any given model’s context limit). This design is suboptimal because there are good times and bad times to compact:

It is not ideal to compact when you’re in the middle of a complex refactor;

It is better to compact when you are starting a new task or otherwise believe that prior context will lose relevance.

Many interactive coding tools feature a /compact command or similar, which allows users to manually trigger a context compression step at opportune times. We take this one step further in the latest release of deepagentsand expose a tool to the agent that lets it trigger context compression itself. This enables more opportunistic compaction without requiring your application’s users to be aware of finite context windows or issue specific commands.

This tool is currently enabled in Deep Agents CLI and opt-in in the deepagents SDK.

We are generally bullish on the idea that harnesses should, where possible, “get out of the way” and take advantage of improvements in the underlying reasoning models. This is an instance of the bitter lesson: can we give agents more control over their own context to avoid tuning their harness by hand?

When should we compact?

There is a variety of situations that could warrant a context compression action.

At clean task boundaries:

A user signals that they are moving on to a new task for which earlier context is likely irrelevant

The agent has finished a deliverable and the user acknowledges task completion

After extracting a result from a large amount of context:

The agent has obtained a fact, conclusion, summary, or other result by consuming a significant amount of context, as in a research task

Before consuming a large amount of new context:

The agent is about to generate a long draft

The agent is about to read a large amount of new context

Before entering a complex multi-step process:

The agent is about to start a lengthy refactor, migration, multi-file edit, or incident response

The agent has produced a plan and is about to begin executing the steps

A decision has been made that supersedes prior context:

New requirements have come to light that invalidate previous context

There are many tangents or dead-ends that can be reduced to a summary

Enumerating all possible scenarios is not practical, but our observation is that people and LLMs can identify these scenarios and compact at opportune times, saving the need for a compaction step later on when the context window is nearing its limit. You can read the guidance we provide the model around this tool in its system prompt.

What happens when the tool is called?

The tool is parametrized the same as the existing Deep Agents summarization middleware: we retain recent messages (10% of available context) and summarize what comes before. Recent messages, including the call to the compaction tool and associated response, are retained in the recent context.

imageSee example trace.

How to use

The tool is implemented as a separate middleware, so you can enable it by adding it to the middleware list in create_deep_agent:

from deepagents import create_deep_agent

from deepagents.backends import StateBackend

from deepagents.middleware.summarization import (

create_summarization_tool_middleware,

)

backend = StateBackend # if using default backend

model = "openai:gpt-5.4"

agent = create_deep_agent(

model=model,

middleware=[

create_summarization_tool_middleware(model, backend),

],

)

See the SDK docs for more details.

In the CLI, simply call /compact when you’re ready to trim context or move onto a new task.

Our experience with this feature

We tuned this feature to be conservative. Deep Agents preserves all conversation history in its virtual filesystem, allowing for recovery of context post-summarization, but an erroneous context compression step is disruptive. We tested:

A custom evaluation suite, in which we used (our own) LangSmith traces to inject follow-up prompts to threads that do and do not warrant compaction;

Terminal-bench-2, in which we did not observe any instances of autonomous compaction;

Our own coding tasks in Deep Agents CLI.

In practice agents are conservative about triggering compaction, but when they do they tend to choose moments where it clearly improves the workflow.

Autonomous context compression is a small feature, but it points at a broader direction for agent design: giving models more control over their own working memory and fewer rigid, hand-tuned rules in the harness. If you’re building long-running or interactive agents, try it out in the Deep Agents SDK or CLI and let us know your feedback and what patterns you’d like to see it handle next.