LangChain: Deep Agents fix AI retrieval waste with files, subagents

LangChain's post frames agent retrieval as a problem of limited context windows, similar to a student taking an open-book exam who has all the answers but struggles to find the right one efficiently. The company uses a competitive analysis agent as a running example: every Monday it researches hundreds of competitors, needs last week's notes on Company A, a specific revenue number from a huge market data export, and anything new in a news archive. Each of these retrieval tasks fails in a different way, and LangChain maps each failure to a technique in its Deep Agents harness.

The first failure, inefficient searching, is addressed by built-in filesystem tools like glob and grep that check names and keywords before opening files, and by adding a semantic retrieval tool when exact keywords are not enough. LangChain notes that grep would miss an article titled "Company A closes $50M Series B" if the agent searched for "Company A funding," which is why semantic search backed by a vector store is needed. The second failure, too much context at once, is handled by reading large files one page at a time with read_file's offset and limit, and by offloading oversized tool results to files. The third failure, irrelevant or stale context, is mitigated by delegating reading to subagents that return only key facts, pulling from up-to-date sources like search APIs, and validating outputs against cited sources.

The post also warns that retrieved text is data, not instructions, and can contain prompt injections that no prompt or delimiter strategy fully prevents, so answers should be checked against expected sources. LangChain summarizes the strategies as: use the most efficient tool that works, page instead of loading, return pointers instead of payloads, delegate the reading, and verify grounding.