Glossary

Standard terminology used throughout MassGen documentation.

.massgen Directory

Organizational directory created by MassGen to store all working files: sessions, workspaces, snapshots, temporary files, and logs.

Agent

An AI assistant configured with a specific backend (model) and capabilities. Agents work independently or collaborate with other agents to solve tasks.

Attempt

A single orchestration execution within a Turn. If orchestration fails or restarts (due to errors, timeouts, or explicit restart requests), each restart creates a new attempt. Attempts are numbered sequentially (attempt_1, attempt_2, etc.). Most turns complete in a single attempt. See also: Turn, Round.

Backend

The AI model provider integration (e.g., OpenAI, Claude, Gemini). Each backend connects to a specific AI service and provides access to its models.

Claude Code

Anthropic’s development-focused AI assistant with native file operation tools (Read, Write, Edit, Bash, Grep, Glob). Backend type: claude_code.

Codex

OpenAI’s Codex CLI with native shell, file editing, web search, and workspace-scoped MCP integration. Backend type: codex.

Consensus

Agreement among agents on the best solution. Agents reach consensus through voting and refinement during coordination phases.

Context Path

A shared directory that agents can access during multi-agent collaboration. Context paths have permissions (read or write) and enable project integration.

Coordination Phase

The period when agents observe each other’s solutions and vote on the best approach. During coordination, agents can refine their answers based on others’ work. Coordination completes when all agents have voted.

File Operation

Reading, writing, editing, or deleting files in an agent’s workspace or context paths. File operations require appropriate permissions.

FileOperationTracker

System component that enforces read-before-delete policies and tracks all file operations per agent for safety.

Final Agent

The agent whose solution is selected after coordination completes. This agent presents the final answer and can execute write operations to context paths.

LM Studio

Local model inference tool that enables running open-weight models on your machine. Backend type: lmstudio.

MassGen

Multi-Agent Scaling System for GenAI. A framework for running multiple AI agents collaboratively to solve complex tasks.

MCP (Model Context Protocol)

A protocol for integrating external tools and services with AI models. MCP servers provide tools like web search, filesystem access, or API integrations.

MCP Planning Mode

Coordination strategy where agents plan tool usage without execution during coordination. Only the final agent executes MCP tools to prevent duplicate actions.

MCP Server

A service that implements the Model Context Protocol and provides tools to agents. Examples: weather service, filesystem access, web search.

Multi-Turn Mode

Interactive conversation mode where you can continue asking questions across multiple turns. Session history is preserved in .massgen/sessions/.

Orchestrator

System component that manages agent coordination, voting, workspace management, and session handling.

Planning Mode

See MCP Planning Mode.

Round

A single LLM call cycle for an agent during coordination, where the agent receives the current context (previous answers), processes it, and produces output (either a new_answer or a vote). Multiple rounds occur as agents iteratively refine solutions. Each round includes: receiving context → LLM streaming → tool execution (if any) → output. See also: Turn, Attempt.

Session

A multi-turn conversation saved in .massgen/sessions/. Sessions preserve context across multiple turns with the same agent team. A session can span multiple user interactions over time.

Snapshot

A workspace state captured during coordination. Snapshots allow agents to share file-based work with each other.

Snapshot Storage

Directory where workspace snapshots are stored. Configured via orchestrator.snapshot_storage in YAML.

Stream Chunk

A single piece of data from a streaming LLM response. Chunk types include: content (text), tool_calls (function invocations), reasoning (thinking), and done (completion signal).

Streaming

Real-time delivery of LLM responses as they are generated, rather than waiting for the complete response. MassGen uses streaming for all LLM calls, enabling live progress display and early tool execution detection.

System Message

Instructions or role description provided to an agent to customize its behavior. Also called “system prompt.”

Temporary Workspace

Directory storing previous turn results for context in multi-turn sessions. Configured via orchestrator.agent_temporary_workspace.

Turn

A single user interaction in a Session. Each turn represents one question/task submitted to the agents, triggering a complete coordination cycle. In multi-turn mode, turns are numbered sequentially (turn_1, turn_2, etc.). Each turn may have one or more attempts. The log directory structure is: log_TIMESTAMP/turn_N/attempt_N/. See also: Round, Attempt.

Voting

Process where agents evaluate solutions and cast votes for the best answer during coordination rounds.

Voting Config

Orchestrator settings controlling voting behavior, including consensus threshold and voting rules.

Voting Threshold

Minimum percentage of agent votes required to select a final answer. Configured as 0.0-1.0 (e.g., 0.6 = 60%).

Workspace

Isolated directory where an agent performs file operations. Each agent has its own workspace under .massgen/workspaces/.

Workspace Isolation

Security feature ensuring each agent’s file operations are contained in their own workspace, preventing accidental interference.

YAML Configuration

Human-readable configuration file format used to define agents, backends, and orchestrator settings in MassGen.

Deprecated Terms

The following terms were used in earlier versions but are no longer preferred:

External Tool

Deprecated when referring to MCP. Use MCP Server or MCP (Model Context Protocol) instead.

Final Presentation Agent

Deprecated. Use Final Agent instead. Simpler terminology preferred.

MCP Tool

Ambiguous. Clarify whether you mean an MCP Server (the service) or a specific tool provided by an MCP server.

Provider

Deprecated. Use Backend instead. Backend is more precise as it refers to the integration code, not just the service provider.

Shared Directory

Deprecated. Use Context Path instead. Context paths are the current terminology for shared project directories.

Winning Agent

Deprecated. Use Final Agent instead. This is the agent selected after coordination completes.

Working Directory

Deprecated. Use Workspace instead. The cwd parameter still refers to the workspace directory.

Configuration-Related Terms

agent

Alternative top-level YAML key (singular) for single-agent configurations. Less commonly used.

agents

Top-level YAML key (plural) for multi-agent configurations. Contains a list of agent definitions.

backend.cwd

YAML parameter specifying the agent’s workspace directory.

backend.model

YAML parameter specifying the model name (e.g., gpt-5-nano, claude-sonnet-4).

backend.type

YAML parameter specifying the backend type (e.g., openai, claude, gemini).

orchestrator.agent_temporary_workspace

YAML parameter for temporary workspace directory location.

orchestrator.context_paths

YAML list of shared directories with read/write permissions.

orchestrator.coordination

YAML section for coordination configuration, including planning mode settings.

orchestrator.snapshot_storage

YAML parameter for snapshot directory location.

permission

YAML parameter for context path access level: "read" or "write".

Backend-Specific Terms

AG2

External framework for agentic workflows with code execution. Backend type: ag2.

Azure OpenAI

Microsoft Azure’s OpenAI deployment service. Backend type: azure_openai.

Chat Completions API

Generic OpenAI-compatible API format. Backend type: chatcompletion.

Codex CLI

OpenAI’s native-tool coding assistant. Backend type: codex.

Gemini

Google’s AI models (Flash, Pro). Backend type: gemini.

GPT

OpenAI’s language models (GPT-4, GPT-5, etc.). Backend type: openai.

Grok

xAI’s language models (Grok-3, Grok-4). Backend type: grok.

Response API

Anthropic’s API format. Backend type: claude.

SGLang

Structured Generation Language for high-performance LLM inference. Supported via chatcompletion backend.

vLLM

High-performance LLM inference engine. Supported via chatcompletion backend.

ZhipuAI / ZAI

Chinese AI provider with GLM models. Backend type: zai.

Usage Examples

Correct terminology in documentation:

The :term:`final agent` executes write operations to :term:`context paths<Context Path>`.

Agents use :term:`MCP servers<MCP Server>` to access external tools.

Each agent has an isolated :term:`workspace` under ``.massgen/workspaces/``.

Preferred vs. deprecated:

Preferred	Deprecated	Context
Context path	Shared directory	Project integration
Workspace	Working directory, cwd	Agent file operations
Coordination phase	Coordination round	When emphasizing the phase, not iteration
Final agent	Winning agent, final presentation agent	Agent selection
Backend	Provider	Model integration

See Also

• Core Concepts - Core MassGen concepts explained

• YAML Configuration Reference - Complete YAML parameter reference

• Configuration - Configuration guide with examples

• CLI Reference - Command-line interface reference