Progressive Decomposition Pattern

Content

{
  "problem": "Agents given large, ambiguous tasks fail by either attempting the full task monolithically (losing coherence) or decomposing too finely upfront (wasting compute on irrelevant subtasks).",
  "examples": [
    "Long-horizon research tasks where early findings change later research direction",
    "Code refactoring where initial analysis reveals architecture different than assumed",
    "Multi-document synthesis where first few documents reframe the question"
  ],
  "solution": "Decompose tasks progressively in layers: first a coarse 3–5 chunk breakdown, then refine only the chunk currently in execution into atomic steps, deferring decomposition of future chunks until their turn.",
  "anti_patterns": [
    "Full upfront decomposition for tasks with high information-dependency",
    "Skipping phase re-evaluation after completing each layer 1 execution",
    "Treating decomposition output as fixed plan rather than living structure"
  ],
  "implementation_steps": [
    "Layer 0: Receive task, generate 3–5 high-level phases. Store as ordered queue.",
    "Layer 1: Dequeue next phase, decompose into 3–7 atomic steps. Do not decompose future phases yet.",
    "Layer 2: Execute atomic steps sequentially. After each step, update shared state with results.",
    "Layer 3: After each phase completes, re-evaluate remaining phases against actual results — revise or discard phases that are no longer needed.",
    "Terminate when phase queue is empty or goal state achieved."
  ]
}