---
id: "contrarian-more-context-is-worse"
type: "contrarian-insight"
source_timestamps: ["00:23:23"]
tags: ["prompt-engineering", "ai-agents", "contrarian"]
related: ["concept-agent-context-scoping"]
challenges: "The conventional view that providing an LLM with maximum possible context yields the best results."
sources: ["s45-claude-limit-chatgpt-habit"]
sourceVaultSlug: "s45-claude-limit-chatgpt-habit"
originDay: 45
---
# Contrarian: More Context Often Degrades Agent Performance

## The Assumption Being Challenged
A common assumption among prompt engineers and agent developers: **more context = better answers**. Pass the whole codebase, the whole manual, the whole knowledge base — let the model figure out what is relevant.

## Nate's Counter-Position
The opposite is closer to the truth in practice. Dumping massive unscoped context into an agent's window:
- Dilutes its attention mechanism
- Degrades its reasoning on the actual task
- And of course wastes tokens (see [[concept-token-burning]])

Context must be **strictly minimized and scoped** — see [[concept-agent-context-scoping]] and the discipline encoded in [[framework-kiss-commands]].

## Supporting Literature (from enrichment overlay)
- **'Lost in the Middle'** (TMLR 2024) — retrieval accuracy drops ~50% in mid-context for long inputs.
- **'Attention-Driven Reasoning'** (arXiv 2403.14932) — non-semantic tokens skew attention; rebalancing yields 10–20% gains without retraining.

## Honest Counter-Counter
- **Needle-in-haystack** tests show some 128K+ context windows remain stable for retrieval-style queries.
- Over-summarizing can lose nuance — 'Long Context RAG' work shows up to ~15% recall loss in some setups.
- The honest middle: **context should be minimized but not amputated**. Pre-process, summarize, retrieve — don't blindly truncate.

## Practical Recipe
Apply [[framework-kiss-commands]]: index references, pre-process, cache stable, scope minimum, measure burn. Use [[entity-claude-code-d45]]'s `/context` command and [[action-measure-context]] to verify you're hitting only the slice the agent actually needs.