Overview
An architecture that combines multiple LLMs (OpenAI, Anthropic, Google, Meta, and in-house open-weight models — often 15+ models), where generated outputs are validated by a Supervisor Agent (a separate model acting as a verifier).
Architecture
-
Worker Layer
- Multiple LLMs generate outputs in parallel or selectively
- Each model is leveraged for its strengths (reasoning, coding, multilingual, etc.)
-
Supervisor Layer
- Validates, compares, and reranks outputs
- Performs hallucination detection, consistency checks, and policy validation
- Optionally triggers regeneration or refinement loops
Characteristics
- Higher reliability and output quality than single-model systems
- Reduced vendor lock-in through model diversity
- Increased cost and latency due to multi-model execution
Use Cases
- High-reliability agents (customer support, finance, legal)
- Code generation with verification
- Stabilizing long-form reasoning tasks
Related Patterns
- Self-Consistency
- Debate / Multi-agent critique
- Reflection / ReAct with verifier
Notes
This pattern can be seen as a combination of:
- Mixture of Experts (MoE)
- Ensemble methods with a critic/verifier model
In Graflow, this maps naturally to:
- Task Graph with fan-out to multiple workers
- Aggregation + Supervisor node for validation
Overview
An architecture that combines multiple LLMs (OpenAI, Anthropic, Google, Meta, and in-house open-weight models — often 15+ models), where generated outputs are validated by a Supervisor Agent (a separate model acting as a verifier).
Architecture
Worker Layer
Supervisor Layer
Characteristics
Use Cases
Related Patterns
Notes
This pattern can be seen as a combination of:
In Graflow, this maps naturally to: