Deep Agents moves long-run state out of the window

LangChain published a conceptual guide on October 14, 2026, on how its open-source Deep Agents harness handles long runs. The post, by Jess Weng, is labeled a conceptual guide. It explains behavior already in the harness. It is not a new model, and it is not a harness release. The same day, the LangChain account asked how agents remember and said longer runs take in instructions, tool results, files, and drafts, which cuts accuracy and reasoning. The account points readers to the harness moving that context out of the active window.

Weng calls the failure context rot. As history, noise, and tool output pile up past what a model's context window can comfortably hold, accuracy and reasoning get worse. Deep Agents does not delete that state. It moves context out of the active window into files, summaries, or a separate agent, so the main agent carries only what it needs and can recover the raw data later. The guide implements each technique as middleware on the agent loop, and says most of it is built in unless you want to customize a step.

Offloading. FilesystemMiddleware runs every time a tool is called. When a result exceeds roughly 20,000 tokens, the full result is written to /large_tool_results/{tool_call_id} before it enters the conversation. The model sees a short preview — the first and last few lines — plus the path, and can page the file later with read_file. Oversized human messages, such as a pasted transcript, use a higher threshold of 50,000 tokens and are saved under /conversation_history/. Those figures are how the guide describes current defaults. It also says thresholds that suit one task rarely suit every task.

Summarization. SummarizationMiddleware checks usage before each model call. Once usage crosses 85 percent of the model's context window, the default, a separate LLM call condenses older messages. Recent messages stay verbatim. The summarized messages are written to /conversation_history/{session_id}.md, so a later step can read a detail back. As the run continues, the next summary can include the previous one. You can also summarize when the topic changes, not only at a token threshold.

Skills. A skill is a folder of instructions, with optional scripts and reference docs. SkillsMiddleware reads only the header of each SKILL.md before the agent starts, then adds those names and descriptions to the system prompt on every model call. Full instructions enter only when the model decides a skill is relevant and reads the file. Weng calls that progressive disclosure: ten skills cost about as much context as ten one-liners until one is needed.

Subagents. For work you already know will be noisy — the guide's example is researching one company across many searches and pages — a subagent runs in its own context window and returns only the result. In the sample, a market-researcher must return three short bullets and must not return the raw report. The post says well-scoped subagent tasks can often run on a cheaper model and in parallel.

Long-term memory. Memory is a markdown file in storage you control. MemoryMiddleware loads it at the start of a run, injects it into the system prompt on every model call, and updates the file when the agent learns something worth keeping, such as a preference for bullets over paragraphs. Because that file is not held by a provider, the same memory file and backend can be used with a different model.

The post says large inputs should be offloaded rather than truncated so they stay retrievable, and that after summarization the agent can still get the detail back.

Read the October 14 guide for the middleware hooks and its competitive-analysis walkthrough. To try the harness the post shows, install with uv add deepagents. Weng also says she will cover this with a live instructor on November 3.

Sources