Anthropic shifts agent engineering toward context engineering

According to @thisdudelikesAI’s breakdown of Anthropic’s official agent engineering guide, Anthropic argues that the bottleneck in long-context, tool-using agents is no longer mainly prompt wording, but how context is stored, filtered, organized, and reused. That makes “context engineering” the practical center of modern agent building, especially for multi-step tasks that accumulate large amounts of history.

Core practices

A central recommendation is to stop stuffing all information into chat history and instead give the agent a persistent, structured notebook. In @thisdudelikesAI’s summary, Anthropic suggests having Claude maintain something like a continuously updated NOTES.md to record progress, decisions, and plans across multiple tool calls. The point is to preserve key conclusions in a reusable form rather than repeatedly replay raw history.

Another major theme is retrieval over total recall. @thisdudelikesAI says the guide cites Anthropic research on “context rot”: as context grows longer, the model becomes worse at locating the truly important details. The recommended response is to persist plans and decisions, then retrieve relevant material when needed, instead of relying on an ever-expanding context window.

Performance and cost

The guide also highlights context editing. Tool outputs that sink deep into the conversation can become what @thisdudelikesAI describes as attention-hurting dead weight, so Anthropic recommends cleaning them up. In Anthropic’s internal agentic search evaluation, combining memory with context editing reportedly produced a 39% improvement.

On the cost side, Anthropic advises freezing the prefix: keep the system prompt and stable project context fixed at the front and identical across turns, while appending changing step-specific information later. The purpose, as summarized by @thisdudelikesAI, is to improve prompt caching hit rates and reduce repeated inference cost.

2026-07-16 ~ 2026-07-16 · 5 related posts