Wire

Six agentic retrieval techniques backed by 2026 benchmarks: read enforcement, query decomposition, routing, split routers, verification, and step budgets.

Pruning an agent's tool history raised task completion from 71% to 91.6%. What the summary rescued was not content but the agent's place in its own work.

Context window blindness: four frontier models misjudged their own context size by 43 to 84%. Why compaction drops the wrong things, and what fixes it.

Studies disagree on the best chunk size for RAG: one found 100 tokens with zero overlap won. What the research supports, and the bigger lever most guides miss.

Ghost memory is when old, current, and changed facts sit in an agent's memory and get retrieved together. Deleting loses history. Labeling the edge works.

A 12,000-trajectory study finds agentic RAG agents finalize answers without reading retrieved evidence. Forcing one read gains up to 19.9 accuracy points.

Across 44 language models, asking for JSON collapsed answer diversity: distinct answers fell from 52 to 36 and the top answer rose from 41% to 64% of responses.

Kimi K3's 1M-token window runs on Kimi Delta Attention: hybrid linear attention with a fixed-size state. Why cheap long context still needs context curation.

GPT-5.6 Sol's Ultra mode gives each subagent its own context window. It solves context isolation but not handoff, and it multiplies your token bill fast.

AI agent reliability fails because the same task assembles different context every run. Non-determinism is a context engineering problem, not a model flaw.