map-reduce

Map-Reduce Pattern

Parallel fan-out to independent mapper agents with merged reduction into a single coherent response.

How It Works

graph LR
    Input --> Splitter
    Splitter --> Mapper1[Mapper 1]
    Splitter --> Mapper2[Mapper 2]
    Splitter --> MapperN[Mapper N]
    Mapper1 --> Reducer
    Mapper2 --> Reducer
    MapperN --> Reducer
    Reducer --> Output

Loading

Flow

1. Split -- The SplitterAgent breaks the input into 2-4 independent sub-tasks using non-streaming provider.chat()
2. Fan-out -- One handoff event per mapper is emitted
3. Map -- All MapperAgents run concurrently via Promise.all(), each streaming its analysis
4. Fan-in -- A single handoff event signals all mappers are done
5. Reduce -- The ReducerAgent synthesizes all mapper outputs into one coherent response

Parallelism

Mappers run in parallel using Promise.all(). Each mapper gets its own LLMProvider instance because lastUsage is instance state on LLMProvider and would race if shared across concurrent streams. Events from different mappers interleave on the stream -- the frontend uses the agent field to attribute chunks to the correct mapper.

Agent Roles

Agent	Role	Streaming	Description
SplitterAgent	splitter	No (chat)	Breaks input into independent sub-tasks
MapperAgent	mapper	Yes (chatStream)	Analyzes a single sub-task in detail
ReducerAgent	reducer	Yes (chatStream)	Synthesizes all analyses into one response