← ClaudeAtlas

concurrent-fanoutlisted

Mechanics of parallel fan-out/fan-in: independence checks, choosing Agent batches (up to 20 calls in one message) vs the /fanout workflow, stable item ids and idempotent briefs, a structured per-item result schema, partial-failure classification with a single retry, collation tables and coverage lines. Load when launching the same operation over many independent items.
rustcircle/c101 · ★ 0 · AI & Automation · score 65
Install: claude install-skill rustcircle/c101
# Concurrent fan-out Fan-out applies one operation to N independent items concurrently and collates the results. It is pure mechanics: no item's result depends on another's, and the collator makes no content judgement (that belongs to a synthesizer, verifier or the caller). If items need partitioning, dedup or a merged answer, that is a swarm (`swarm-execution`). Brief, schema and retry templates ready to copy: [templates](references/templates.md). Collation helper: [collate.py](scripts/collate.py) turns item JSON into the table and coverage line. ## 1. Preconditions - [ ] **Independent**: no item needs another item's output. - [ ] **Disjoint writes**: no two items write the same file. Read-only work is ideal. Shared build directories count as shared state: concurrent `cargo`, `gradle` or `npm` runs in one checkout contend for locks; give each item its own target dir or serialize. - [ ] **One agent type per item** (or an explicit item -> agent type map). - [ ] **Schema** for the per-item result defined before launch (section 4). - [ ] **Budget**: `N + ceil(0.2 N)` (retry reserve) at most `H/2`, with `H = max_spawns_per_session - spawns_this_session` (`delegation-protocol`). - [ ] **Depth**: `remaining_depth >= 1`; at 1-2 the item agent must be a leaf (no coordinators). - [ ] **Token budget**: header `budget="ok"`; at `"soft"` launch only the items the deliverable needs (converge, no exploratory items); at `"hard"` launch no fan-out (only `file-summ