Adaptive vs Fixed Decomposition

Choosing between fixed and adaptive decomposition is a critical architectural decision that affects both quality and efficiency.

When to Use Fixed Pipelines

• The workflow has well-defined stages (e.g., extract -> validate -> transform -> load)
• Each stage has predictable inputs and outputs
• The number of steps is known in advance
• Quality comes from thoroughness at each stage, not exploration

When to Use Adaptive Decomposition

• The scope is unknown until investigation begins
• Early findings change what needs to be investigated next
• Dependencies between subtasks are discovered during execution
• The task requires exploration before planning

Attention Dilution

Large inputs (e.g., reviewing a 50-file PR in one pass) cause attention dilution -- the model may miss issues in the middle of long contexts. The solution is decomposition:

1. Per-file local analysis (each file gets focused attention)
2. Cross-file integration pass (finds interactions between files)

This two-pass approach catches both local issues and cross-cutting concerns.