LLM Inference Interview Questions, issue 2, Jul 31, 2026
The Context Compression Trap
AI Agent Engineer interview at Anthropic, and the interviewer asks:
“Your coding agent keeps overflowing the context window. You add LLM summarization to condense old steps, cost drops 2x on SWE-bench. But now the agent opens the same pull request three times. What did summarization actually break, and how do you fix it without losing the savings?”
Don’t say: “The summary lost some detail, so I’ll write a better summarization prompt.”
Why condensing your agent's history quietly causes infinite action loops, and the architectural fix to separate lossy reasoning from irreversible state mutations.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.