Skip to content

Chain-of-Experts Pattern

Sequential expert processing pipeline where each specialist builds on the last.

The Chain-of-Experts pattern passes a task through a chain of specialized expert agents, each adding their perspective before passing to the next expert. A final synthesizer integrates all contributions into a coherent output.

This pattern excels at complex tasks requiring multiple domain perspectives, where each expert's analysis builds upon and informs the next.

When to Use

Good fit Poor fit
Complex research requiring multiple domain expert reviews Simple, single-domain tasks
Document editing with sequential specialist passes Tasks requiring parallel processing for speed
Multi-stage analysis where each expert builds on the previous Tasks with a single correct answer
Legal, medical, or technical review pipelines Situations needing immediate response

Architecture

flowchart LR
    A([START]) --> B["expert_1\n— First expert analysis —"]
    B --> C["expert_2\n— Second expert builds on first —"]
    C --> D["expert_N\n— Nth expert contribution —"]
    D --> E["synthesizer\n— Final integration —"]
    E --> F([END])

State flows through the graph:

Field Type Description
task str The input task
experts list[dict] List of expert definitions [{name, specialty, system_prompt}]
current_expert_index int Index of the current expert being processed
expert_outputs list[dict] Accumulated outputs from each expert
final_synthesis str Final integrated output from synthesizer

Core Code

from patterns.chain_of_experts.pattern import ChainOfExpertsPattern

pattern = ChainOfExpertsPattern()

experts = [
    {"name": "Legal Reviewer", "specialty": "legal analysis"},
    {"name": "Technical Expert", "specialty": "technical feasibility"},
    {"name": "Risk Analyst", "specialty": "risk assessment"},
]

result = pattern.run(
    task="Review this software license agreement",
    experts=experts,
)

print(result["final_synthesis"])  # Integrated expert analysis

Configuration Options

Parameter Default Description
model "gpt-4o-mini" OpenAI model name (ignored when llm is provided)
llm None Pre-configured LangChain BaseChatModel instance

Quick Start

# 1. Clone and install
git clone https://github.com/your-org/agentflow.git
cd agentflow && uv sync

# 2. Set your API key
echo "OPENAI_API_KEY=sk-..." > .env

# 3. Run the example
uv run python -m patterns.chain_of_experts.example

Example Output

============================================================
CHAIN-OF-EXPERTS PATTERN -- Multi-Expert Analysis
============================================================

Task: Analyze a proposed software partnership agreement

Expert 1: Legal Reviewer
Expert 2: Technical Expert
Expert 3: Risk Analyst

============================================================
FINAL SYNTHESIS:
============================================================
# Comprehensive Partnership Analysis

## Legal Perspective
[The legal reviewer's analysis of contract terms...]

## Technical Feasibility
[The technical expert's assessment...]

## Risk Assessment
[The risk analyst's evaluation...]

## Integrated Conclusion
[Synthesized findings from all three experts...]

============================================================
Expert Contributions: 3

How It Works — Step by Step

  1. Initialization: The graph receives a task and a list of expert definitions.
  2. Expert Processing: Each expert in sequence processes the task, receiving context from all previous expert outputs.
  3. Sequential Pass: Expert 1 → Expert 2 → ... → Expert N, each building on previous contributions.
  4. Synthesis: A final synthesizer node integrates all expert outputs into a coherent conclusion.
  5. Output: The graph returns the complete synthesis reflecting all expert perspectives.

Comparison with Other Patterns

Dimension Chain-of-Experts Reflection Debate
Agent count N experts + 1 synthesizer 2 (writer + reviewer) 2+ adversarial
Interaction Sequential chain Sequential loop Adversarial rounds
Best for Multi-domain analysis Iterative refinement Opposing viewpoints
Output Integrated synthesis Improved draft Win/lose verdict
Complexity Medium Low Medium

Chain-of-Experts is ideal when you need multiple specialized perspectives that build upon each other. Use Reflection when you need iterative improvement on a single output. Use Debate when you need to explore opposing viewpoints.

Running Tests

uv run pytest patterns/chain_of_experts/tests/ -v

Tests use mocked LLMs and require no API key.

File Structure

patterns/chain_of_experts/
├── __init__.py
├── pattern.py        # Core ChainOfExpertsPattern class
├── example.py        # One-click runnable demo
├── diagram.mmd       # Mermaid architecture diagram source
├── README.md         # This file (English)
├── README_zh.md      # Chinese documentation
└── tests/
    ├── __init__.py
    └── test_chain_of_experts.py