Namespace JD.AI.Core.Agents.Orchestration.Strategies

Classes

BlackboardStrategy

Blackboard strategy — agents operate on a shared knowledge base (the scratchpad) in iterative rounds. Each round, agents read the current blackboard state, contribute their analysis, and write results back. Converges when no agent contributes new information or max iterations reached. Good for: complex analysis, multi-perspective investigation, knowledge synthesis.

DebateStrategy

Debate strategy — multiple agents independently answer the same question with different perspectives (e.g., optimist, skeptic, pragmatist), then a judge agent evaluates and synthesizes the best answer. Good for: architectural decisions, risk assessment, trade-off analysis.

FanOutStrategy

Fan-out / fan-in — all agents run in parallel, then a synthesizer agent merges their results into a single coherent output. Good for: parallel code review, multi-perspective analysis.

MapReduceStrategy

Map-Reduce strategy — split input into chunks, process in parallel, then reduce. Good for: bulk file analysis, large codebase scanning, data processing.

PipelineStrategy

Pipeline strategy — a typed sequential execution where each agent's output is explicitly transformed before being passed to the next stage. Unlike Sequential/Relay which pass raw output, Pipeline stages have defined input/output contracts and can specify transform functions. Good for: CI/CD workflows, ETL processes, staged code review.

RelayStrategy

Relay strategy — agents process sequentially, each refining the previous output. Like a relay race where each runner adds their specialized perspective. Good for: document writing, iterative improvement, multi-perspective analysis.

SequentialStrategy

Sequential pipeline — agents run in order, each receiving the previous agent's output appended to its prompt plus access to the shared scratchpad. Good for: explore → plan → implement → review pipelines.

SupervisorStrategy

Supervisor strategy — a coordinator agent dispatches work to worker agents, reviews their results, and can redirect or retry until satisfied. Good for: complex multi-step tasks where quality matters.

VotingStrategy

Voting strategy — N agents independently process the same input. Results are aggregated by configurable voting (majority, weighted confidence). Good for: code review consensus, classification, risk assessment.

Enums

VotingMethod

Method for aggregating votes.