atomicity-criteria

name	atomicity-criteria
description	Criteria for judging if a task is atomic (executable in one step). Use with CheckAtomicity to decide when to Decompose vs FormatTask. Essential for atomicity decisions. Apply when processing each task in the planning workflow.

Atomicity Criteria

A task is atomic (atomic: true) if ALL of the following hold. Otherwise, call Decompose.

Four Criteria

1. Single-step executable: Can be completed in one LLM/expert call or one focused work session.
2. Clear deliverable: Has a well-defined output (document, list, report, artifact). No ambiguity about "done".
3. No sub-phases: Cannot be meaningfully split into distinct steps with different responsibilities.
4. Reasonable granularity: Appropriate scope for one session—neither trivial ("open a browser") nor too broad ("complete the entire research").

Decision Flow

Can you name 2+ distinct sub-phases with different deliverables?
├── YES → atomic: false (Decompose)
└── NO → Is there one clear output in one session?
         ├── YES → atomic: true (FormatTask)
         └── UNCERTAIN → Prefer atomic if deliverable is well-defined

Decision Rule

• atomic: false → You must be able to list 2+ specific sub-phases with different deliverables. Example: "Phase 1: collect data produces A; Phase 2: analyze produces B."
• atomic: true → One clear output, one focused session. When in doubt between one vs multiple sessions, prefer atomic if the deliverable is well-defined.

Examples: Atomic

Task	Why atomic
"确定文献检索关键词与数据库范围"	One scoping decision, one output: keywords + scope
"检索并筛选2023-2024年AI领域核心论文"	One search session, one deliverable: filtered list
"撰写文献综述报告"	One writing session, one document
"定义假设与评估指标"	One decision, one output: hypothesis + metrics
"整理要点并撰写 Python 后端调研报告"	One synthesis session, one report (input already gathered)
"对比两者优缺点并撰写总结报告"	One comparison session, one report (inputs from prior tasks)

Examples: Non-Atomic

Task	Why non-atomic
"文献调研：系统收集和综述..."	Explicit phases: collect + review, different deliverables
"实现并测试数据预处理模块"	Implementation vs testing are distinct phases
"调研Python与JavaScript生态并撰写对比报告"	Research phase + synthesis phase
"设计并执行实验"	Design vs execute are distinct
"调研 Python 后端生态：框架、性能、适用场景"	Scope → Gather → Synthesize are distinct phases
"收集资料并分析趋势"	Gather vs analyze are distinct methodologies

Borderline Cases

• Single search + single filter: Often atomic if one session produces one list.
• Multiple methodologies: If "research A" and "research B" are parallel, they can be separate atomic tasks. If "research then analyze" are sequential, decompose.
• Deep hierarchy: At deeper levels, tasks are often already more specific—lean toward atomic when deliverable is clear.
• "整理并撰写" vs "收集并撰写": "整理并撰写" (organize + write from existing input) → often atomic. "收集并撰写" (gather + write) → non-atomic.
• Context from siblings: If a task's input comes from prior sibling artifacts, it may be atomic (e.g. "对比并撰写报告" when 1 and 2 produced research outputs).