Map-Reduce
Fan-out across data partitions in parallel; a reduce node aggregates the outputs. Built-in parallelism.
A map-reduce agent splits a job across partitions, processes each in parallel, then aggregates the partial results in a reduce step. It is the classic parallel shape: the same operation applied independently to many pieces, then combined, so throughput scales with the number of workers. It suits embarrassingly-parallel work over large inputs, where the pieces do not depend on each other. In anatomy terms it is a coordinator that fans work out to identical scoped cores and folds their outputs back together.
A coordinator maps the input across N identical worker cores in parallel, then a reduce core aggregates their outputs; workers are independent and hold no shared state; the perimeter gates the coordinator's writes.
- ·Processing large document or data sets in parallel.
- ·Fan-out extraction or summarisation over many sources, then merge.
- ·Any embarrassingly-parallel workload with an aggregation step.
- ·Throughput scales with the worker count.
- ·Simple, well-understood shape.
- ·Workers are independent: no cross-talk.
- ·Bounded lifecycle: map, reduce, done.
- ·Only fits work that partitions cleanly.
- ·The reduce step can become the bottleneck.
- ·Stragglers slow the whole job.
- ·No cross-partition reasoning during the map phase.
- Internal research (Scalarity), Agents vs Operators / Agent Archetypes v2 · unvalidated
Description, use cases and trade-offs: Field Book analysis, synthesised from established agent-architecture literature.