Summary
At the AI Dev 25 NYC conference, Nyah Macklin—senior engineer and developer advocate whose research spans graph theory, graph algorithms, and context engineering—presented a practical framework for giving AI agents the right information at the right time in the right format. The talk uses a high-stakes security incident scenario (production is down, attacker moving through the network) to show exactly why generic agents fail: without structured, relevant context, even capable LLMs return generic checklists instead of system-specific, actionable steps.
Macklin walks through several concrete context engineering techniques: context pruning to remove irrelevant information before it reaches the model, structured ordering following Anthropic’s documented guidance to place high-priority information at the top of the context window, and tool offloading—giving agents database query tools rather than dumping raw tables into the prompt. The centerpiece is knowledge graph-augmented context, which encodes relational facts as a traversable graph, enabling multi-hop reasoning that vector similarity search routinely misses. Macklin contrasts graph and embedding representations directly, showing why “A is related to B via X” context outperforms nearest-neighbor retrieval for complex queries.
The session closes with pointers to relevant white papers and courses, making it a useful entry point for engineers moving from prompt engineering toward full context system design for production agents.
📺 Source: DeepLearningAI · Published December 08, 2025
🏷️ Format: Deep Dive
