Skill · pattern · parallelization
Fan-out / reduce
Run independent subtasks in parallel and reduce in code (sectioning or voting). Use when subtasks do not share mutable state: multi-file review, scatter-gather research, or LLM-as-judge voting.
GET /api/canon/skills/parallelization?format=md
LangGraph Send, map-reduce, and Anthropic's parallelization catalog. Code aggregates. The model does not 'merge vibes'.
When
Independence is real. Shared mutable state makes this a race, not a pattern.
Do
- 01
Split in code
Sections, files, or candidates. Each worker gets a sliced window.
- 02
Reduce in code
Concat, vote, or structured merge. Cap fan-out.
- 03
No shared writes
Workers write to distinct paths. Parent synthesizes.
Don't
- Fan-out onto the same ticket row without isolation.
- Unbounded map over 'all the files'.
Hard rules
- Independent → parallel. Discovered at runtime → orchestrator-workers.
- Cap breadth the way you cap steps.
Refuse
- Shared mega-context — A crew sharing one bloated window. Specialists drown in each other's work.
- Unbounded loop — No max-turns, token, wall-clock, or money budget. The agent rambles until the bill or the context dies.
Load with this
Load next
Trigger tests
Should fire
- “Review these 12 files in parallel”
- “Vote with three judges”
Should not
- “The steps must run in order”