Latent Space: The AI Engineer Podcast Context Engineering for Agents - Lance Martin, LangChain
1034 snips
Sep 11, 2025 Lance Martin, a key figure at Langchain specializing in context engineering, dives deep into effective tools for AI agents. He discusses the critical importance of context management during tool interactions and differentiates it from traditional prompt engineering. The conversation highlights innovative strategies to improve information flow, reduce context loss, and tackle challenges like context poisoning. Lance also sheds light on the evolution of context engineering, exploring how new buzzwords and practices shape the future of AI technology.
AI Snips
Chapters
Transcript
Episode notes
Context Engineering Emerged From Agent Complexity
- Context engineering arose because agents receive context from many tool calls, not just user prompts.
- Managing that flowing context is the central new challenge when building agents.
Offload Raw Tool Outputs
- Offload heavy tool-call outputs to external storage instead of stuffing them into the message history.
- Store raw results on disk or in agent state and send compact summaries or URLs back to the model on demand.
Summarize For High Recall
- Use careful prompted summarization to compress documents with high recall for later retrieval.
- Tune prompts or fine-tune models to produce exhaustive bullet summaries so the agent knows when to fetch full context.