CORE GUIDE
Workflow decomposition
Break a complex goal into stages with explicit inputs, outputs and verification before adding more agent intelligence.
Mental model
A useful stage has one clear responsibility, a bounded contract and a handoff another stage can inspect without sharing the entire scratchpad.
Why it matters
Large prompts mix research, planning, execution, checking and publishing into one opaque step. Decomposition makes failures attributable, retries targeted and deterministic components easier to substitute for model reasoning where they are better suited.
01
Decompose by responsibility
Identify transformations such as gather evidence, decide, act, verify and package. Give each stage explicit state and postconditions. Introduce parallelism only when work is genuinely independent or when independent verification provides more value than the coordination cost.
02
Example: source-backed article production
A robust workflow separates source collection, claim extraction, outline design, drafting and independent verification. The draft receives a claim-and-source artifact rather than every research scratch note, so a factual defect can be traced to a specific stage.
Common failure modes
- Creating many agent personas without distinct responsibility boundaries.
- Sharing one growing scratchpad across every stage.
- Retrying the entire workflow when only one stage failed.
Engineering heuristics
- Name stages after artifacts or decisions, not personas.
- Keep handoffs small and inspectable.
- Prefer a simple sequential topology until concurrency produces measurable value.
Takeaways
- 01Decomposition improves observability before it improves sophistication.
- 02A stage earns its existence when it localizes responsibility.
- 03Complex topology must earn its coordination cost.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.