Core Resources

MesaLogo platform's core resources are divided into template layer and instance layer, adopting a template-instance separation design pattern.

Resource Overview

Resource	Layer	Description
ActionSpace	Template	Scenario template, defines role combinations, rules, and shared variables
Role	Template	Agent template, defines system_prompt, capabilities, knowledge base
Capability	Template	Agent capability definition (e.g., memory, planner)
RuleSet	Template	Behavior rules and constraints
Knowledge	Template	Knowledge documents accessible by roles
ActionTask	Instance	Scenario instance, actual task run by users
Agent	Instance	Role instance, specific agent in task
Conversation	Instance	Dialogue instance, a task can have multiple sessions
Message	Instance	Dialogue content
EnvironmentVariable	Instance	Task-level variable, visible to all Agents
AgentVariable	Instance	Agent-level variable, visible only to that Agent

Template Layer Resources

ActionSpace

Action space is the template definition of a scenario, containing:

• Basic Information: Name, description, background setting
• Role Configuration: List of participating roles and their configurations
• Rule Sets: Applicable rule collections
• Shared Variables: Public variable templates visible to all roles

{
  "name": "Expert Meeting",
  "description": "Multi-expert collaborative discussion scenario",
  "roles": ["Technical Expert", "Product Manager", "Project Manager"],
  "rule_sets": ["Sequential Speaking Rule", "Time Limit Rule"],
  "shared_variables": {
    "meeting_topic": "string",
    "time_limit": "number"
  }
}

Role

Role is the template definition of an agent, containing:

• System Prompt: Role's system prompt
• Capability Configuration: Capabilities the role possesses
• Knowledge Base: Knowledge bases accessible by the role
• Tool Configuration: MCP tools usable by the role

{
  "name": "Technical Expert",
  "system_prompt": "You are a senior technical expert...",
  "capabilities": ["memory", "knowledge_access"],
  "knowledge_bases": ["Technical Documentation Library"],
  "tools": ["code_executor", "file_reader"]
}

RuleSet

Rule set defines behavior rules and constraints:

• Natural Language Rules: Rules described in natural language
• Logic Rules: Precise rules defined in code
• Trigger Conditions: Conditions for rule activation
• Priority: Priority when rules conflict

{
  "name": "Sequential Speaking Rule",
  "type": "natural_language",
  "content": "Participants speak in order, each person's speaking time should not exceed 5 minutes",
  "priority": 1
}

Knowledge

Knowledge base stores knowledge accessible by roles:

• Type: LightRAG, GraphRAG, Graphiti
• Documents: List of uploaded documents
• Configuration: Retrieval parameter configuration

Instance Layer Resources

ActionTask

Action task is the instantiation of action space:

• Source: Created from ActionSpace
• Agents: List of instantiated Agents
• Sessions: List of dialogues in the task
• Environment Variables: Task-level variable instances

Agent

Agent is the instantiation of role:

• Role Reference: Associated Role template
• Private Variables: Agent-specific variables
• State: Current state information

Conversation

Conversation manages dialogues between agents:

• Participants: Agents participating in the dialogue
• Message List: Dialogue message records
• Autonomous Tasks: Automatic dialogue tasks

Message

Message is the basic unit of dialogue:

• Sender: Agent sending the message
• Content: Message content
• Type: Text, tool call, system message
• Timestamp: Send time

Variable System

EnvironmentVariable

Task-level public variables, visible to all agents:

{
  "project_deadline": "2024-03-01",
  "budget": 100000,
  "priority": "high"
}

AgentVariable

Agent private variables, visible only to that agent:

{
  "expertise_level": 5,
  "available_hours": 40,
  "current_tasks": ["task1", "task2"]
}

Data Flow

User Operation              Template Layer              Instance Layer
──────────────              ──────────────              ──────────────

Select Action Space ──→ ActionSpace
                           │
                           ↓ Create Task
                     ┌─────────────┐
                     │ ActionTask  │ ←─ Copy roles, variables
                     └─────────────┘
                           │
                           ↓ Auto-create
                     ┌─────────────┐
                     │ Conversation│ ←─ Default session
                     └─────────────┘
                           │
Start Autonomous Task ────→ ↓
                     ┌─────────────┐
                     │AutonomousTask│ ←─ Control dialogue rounds
                     └─────────────┘
                           │
                           ↓ Generate
                     ┌─────────────┐
                     │  Message[]  │ ←─ Agent dialogue content
                     └─────────────┘

Parallel Experiments

ParallelExperiment supports controlled variable experiments:

• Source: Binds to ActionSpace scenario template
• Cloned Tasks: Creates multiple ActionTask clones
• Variable Isolation: Each cloned task has independent variables
• Result Comparison: Compares results of different variable combinations

ParallelExperiment
    │
    ├── source_action_space_id → ActionSpace
    │
    └── cloned_action_task_ids → [ActionTask, ActionTask, ...]
                                      │
                                      ↓
                                 Conversation
                                      │
                                      ↓
                                AutonomousTask