Context Management for Deep Agents

As the length of tasks handled by AI agents increases, context management to prevent information loss (Context Rot) within the LLM's limited memory has emerged as a key challenge.

LangChain's open-source agent framework, Deep Agents SDK, provides various Context Compression features for agents performing complex, long-running tasks. This SDK helps agents plan, spawn sub-agents, and execute tasks while interacting with the file system.

Deep Agents executes three main compression techniques in stages depending on the context size:

• Offloading large tool results: When a tool response exceeding 20,000 tokens occurs, its content is moved to the file system, leaving only the file path and a portion of the content (the first 10 lines) in the context.
• Offloading large tool inputs: When the context exceeds 85% of the model's allowance, duplicate tool call arguments (such as file write/edit contents) that have already been saved as files are deleted and replaced with file pointers.
• Summarization: If offloading alone is not enough to free up space, an LLM is used to convert the conversation history into a structured summary that is kept in memory. At this time, the original conversation content is preserved in the file system so it can be retrieved if needed.

Through this multi-pronged approach, agents gain the ability to clearly recognize the goal and progress of a task while still being able to retrieve necessary details at any time.