Pattern Overview

Available reusable pattern implementations:

• TwoSpeakerConversationPattern: two-speaker LLM conversation loops with per-speaker prompts and clients.

• DebatePattern: multi-round affirmative/negative debate with a judging pass.

• PlanExecutePattern: planner + executor decomposition for complex tasks.

• ProposeCriticPattern: fixed two-role proposal/critique loops with approval gating.

• RouterDelegatePattern: intent-based delegation to selected agents.

• RoundBasedCoordinationPattern: round-based peer coordination with no central orchestrator.

• BlackboardPattern: networked coordination with explicit shared board semantics.

• TreeSearchPattern: generator/evaluator reasoning over configurable tree-search strategies.

• RalphLoopPattern: ordered multi-role refinement with dedicated evaluator-score threshold stopping.

• NominalTeamPattern: independent fan-out generation with evaluator-driven best-of-N selection.

• RAGPattern: retrieval-augmented reasoning via memory workflow steps.

Patterns are concrete workflow implementations, not construction primitives. Use design_research_agents.workflow when composing new workflow graphs.

All pattern customization is constructor-first. Helper factory functions were removed in favor of explicit class initialization.

Examples

• examples/patterns/plan_execute.py

• examples/patterns/propose_critic.py

• examples/patterns/router_delegate.py

• examples/patterns/debate_pattern.py

• examples/patterns/two_speaker_conversation.py

• examples/patterns/coordination_patterns.py

• examples/patterns/tree_search.py

• examples/patterns/ralph_loop.py

• examples/patterns/nominal_team.py

• examples/patterns/rag.py