Effective context engineering for AI agents (www.anthropic.com)

After years of prompt engineering, Anthropic and others are promoting "context engineering": the practice of optimizing the entire set of tokens available to an LLM during inference (system prompts, tool outputs, examples, message history, external data, etc.) rather than just crafting single prompts. This matters because transformer-based LLMs have a finite "attention budget" — every token competes in n² pairwise attention, leading to "context rot" where accuracy and long-range reasoning degrade as context grows. Techniques like position-encoding interpolation can extend context length but with degraded position fidelity, so context size trades off against attention focus and precision. Practically, context engineering means curating a minimal, high-signal context: clear, Goldilocks-level system prompts; compact, well-scoped tools with unambiguous parameters; diverse canonical few-shot examples instead of exhaustive rule dumps; and dynamic retrieval strategies. Agentic systems (LLMs using tools in a loop) benefit from "just-in-time" retrieval—storing lightweight identifiers and loading only needed data at runtime—plus metadata (file paths, timestamps, folder structure) and progressive disclosure to keep working memory tight. The payoff is more steerable, robust agents that avoid brittleness and token bloat, at the cost of added engineering for tool design and runtime exploration latency.