How to compress the context when exceeding the token limit

[Smile0524]

How to compress the context when exceeding the token limit in sdk，just like claude code

[Berk Demirci (demirciberk)]

Deep Agents already has built-in context compression, so in most cases you don't need to manually truncate the conversation.

It uses two main mechanisms:

Large tool inputs/results are offloaded to the filesystem and replaced with references.
When the context approaches the model's input limit, older messages are summarized while recent context is preserved.

In the current implementation, summarization typically triggers around 85% of the model's max_input_tokens. If a model call raises a ContextOverflowError, Deep Agents can also summarize the history and retry.

For long-running workflows, subagents are also useful because they isolate detailed tool work and return only the final result to the parent agent.

Docs: https://docs.langchain.com/oss/python/deepagents/context-engineering

[grandocean (173787247)]

Deep Agents does not ship Claude Code’s exact auto-compact, but you can get the same shape in the SDK: summarize old turns + keep a filesystem scratchpad.

Practical recipe

1. Offload, don’t stuff the prompt. You already have FilesystemBackend — write large tool results / retrieved docs to files and put paths + 1-line summaries in the messages. That is the Claude Code pattern (write then read slices).
2. Middleware compact. On context overflow / when token_count > threshold, run a summarizer model over messages[:-N] and replace them with one Human/AI summary message. Keep the last N turns raw. LangGraph + Deep Agents hooks are the place (pre-model hook), not create_deep_agent(...) flags.
3. Cap tool output at the tool wrapper (trim, schema, headroom-style compress) so you never hit the wall in one read_file.
4. Local/open-weight models: they compact worse. Prefer smaller max output + more files; don’t rely on the model to “remember the summary faithfully”.

There isn’t a single compress_context=True equivalent to Claude Code in the SDK today. Compose: backend files + explicit summary node + tool-output caps. If you need the productized version, that’s what hosted Deep Agents / LangSmith tracing is for — locally you wire the hook.

[Sai Charan (saicharan-017)]

1. Filesystem Offloading: Write large tool outputs or file contents to a local scratchpad file and inject only a 1-line summary and path back into the prompt.
2. Threshold Summarization: Use a pre-model middleware hook to monitor token counts (e.g., at ~85% capacity). Summarize older messages (messages[:-N]) into a dense state brief while keeping the last $N$ turns raw.
3. Tool Output Caps: Cap responses at the wrapper level to prevent a single file read from blowing out your context.