When better retrieval makes agents worse
summary
Argues that improving retrieval scores can worsen looped agents because hard distractors—plausible but wrong passages like “2022” vs “2026”—steal attention, causing inverse scaling where longer reasoning produces confident mistakes. Shows the first-drop nonlinear impact: the earliest distractor tokens disproportionately corrupt multi-step plans. Demonstrates a coding-agent experiment on BrowseComp+ raising F1 from ~80 to 88.5, and that adding conventional vector/text search reduced performance. Recommends context admission control: low K, stricter rerank/prune, sufficiency checks, sandboxed primitives for verification.
AI-written summary · about
chapters (10)
- 0:00Introduction to agentic retrieval
- 3:18Traditional retrieval vs agentic retrieval
- 8:56Context injection in agent loops
- 12:54Transformer context limits and accuracy
- 16:05The problem with distractors
- 23:25Rethinking retrieval metrics for agents
- 27:11Precision over recall for agent safety
- 30:01BrowseComp Plus and coding agents
- 33:34Agentic retrieval as context control
- 35:24Q&A on ranking pipelines
related talks
49mContext is King
16mUser Signal Dies at the Retrieval Boundary
43mAdvanced Prompt Engineering and Beyond
15mMemory-Augmented AI Agents: Enhancing Long-Term Reasoning and Context Awareness
11mThe Unreasonable Effectiveness of Prompt Learning
25mAdvanced RAG Systems: Scaling Retrieval-Augmented Generation in Production
01h 30mOptimizing Retrieval for Agentic Systems
26mWhy More Context Makes Your Agent Dumber and What to Do About It