Config

This guide outlines how to define Team Structure, Collaboration Logic, and Evolutionary Strategies for your multi-agent system.

---

Minimal YAML Template

domain: "dev_ops"
num_rounds: 3

agents:
  - name: "Lead"
    role: "System Architect"
    count: 1
  - name: "Dev"
    role: "Python Engineer"
    count: 2

connection_mode: "star"
decision_method: "DecisionMethod.REFER"

llm:
  model_name: "gpt-4o"
  temperature: 0.2

optimization:
  strategy: "prune"
  pruning_rate: 0.2

---

Core Logic

The configuration follows a top-down hierarchy:

Brain (LLM) → Members (Agents) → Organization (Topology) → Evolution (Optimization).

---

Parameter Definitions

LLM Layer:

Defines the cognitive engine for all agents.

Parameter	Range	Recommendation
model_name	str	Use gpt-4o for logic; gpt-4o-mini for cost-efficiency.
temperature	0.0 - 1.0	0.1-0.3: Coding/Logic; 0.7-0.9: Creative writing.
max_tokens	int	Prevents truncation; 2048+ recommended for code generation.

---

Agent Layer:

Defines the specialized roles in your team.

role: The System Prompt. Be specific about expertise and output format.

count: Number of instances. Increase to gain diverse perspectives via ensemble logic.

---

Topology & Decision Defines how information flows and how the final result is reached.

connection_mode:

• full_connected: Everyone sees everything. High transparency, high noise.
• star: First agent is the Leader. Centralized command and control.
• chain: $A \to B \to C$. Ideal for sequential pipelines (Design → Code → Test).
• layered: Mimics organizational hierarchies.

decision_method:

• major_vote: Best for objective tasks (Math/Logic).

• refer: Final summary by a designated expert (usually the last agent).

• weighted: Aggregates results based on agent importance.

---

Optimization Layer Automates the refinement of communication paths over time.

strategy:

• prune: Removes redundant or low-value communication edges.

• dropout: Randomly breaks connections to force agent independence.

• Bayesian: Uses Bayesian optimization with optional MCMC sampling for uncertainty-aware pruning.

pruning_rate: (0.0-1.0). Default 0.2. High values lead to a more sparse, efficient network.

optimize_spatial: Set to true to allow the system to re-route "who talks to whom."

---