Context Engineering
A 1M-token window is a budget, not free space. Curate the smallest high-signal set of tokens on every turn, and keep the prefix cacheable and append-only.
The window fills up and the agent loses the thread
Frontier windows are about 1M tokens (Claude Opus 5.5 and Sonnet 5.5, GPT-6.1 Sol at 1.05M, Gemini 3.8 Flash). An agent fills its window turn by turn with tool output and history, and accuracy drops as the count grows. This is context rot: attention spans n² pairwise relationships, and training data skews toward shorter sequences.
Serving cost rises too. One study (arXiv 2611.05306) found that serving Qwen 3 Coder Next with a 256K window instead of 16K costs 7× more per generated token, while frequent compaction kept frontier agents about as accurate on SWE-bench Verified.
Lesson 5 of 8 in Prompting & Context