MAS Design

Collaborative Multi-Agent Architecture Directory

Repo
...
30 of 30 architectures

Add New Architecture

Create and visualize your multi-agent architecture with our builder.

Simplified view

Collaborative Circle Architecture

A collaborative multi-agent system where agents contribute to a shared memory and iteratively refine a solution through open, cyclical information sharing.

collaborativeshared-memory+2
@eren9677
2025
User Input
input
Agent 1
agent
Agent 2
agent

Basic Sequential

Agents operate in a linear pipeline; each agent’s output is the next agent’s input.

sequentialpipeline+2
@eren9677
2025
Agent 1
agent
Agent 3
agent
Agent 4
agent
Agent 2
agent

Network (Decentralized) Architecture

Agents communicate directly without a central coordinator for flexible, dynamic collaboration.

p2pdecentralized+2
@eren9677
2025
Simplified view

Market-Based (Bidding) Architecture

Auctioneer broadcasts tasks; contractor agents bid; best bid wins based on predefined criteria.

biddingauction+3
@eren9677
2025
Agent Supervisor
supervisor
Sub-Supervisor 2
supervisor
Sub-Supervisor 1
supervisor
Agent 1a
agent
Agent 2b
agent
Agent 2a
agent
Agent 1b
agent

Hierarchical Architecture

Multi-level supervision where top-level supervisors manage lower-level supervisors who manage worker agents.

hierarchicalsupervisor+2
@eren9677
2025
User Goal
input
Orchestrator Agent
agent
Web Search API
tool
Code Interpreter
tool
Database
tool
Final Result
output

Tool-Augmented Agent Network

Agents are designed to interact with external tools (APIs, databases, code interpreters) to ground reasoning and perform real-world actions.

toolsapis+4
@eren9677
2025
Initial Task
input
Creator Agent
agent
Critic Agent
agent
Final Output
output

Critic & Refinement Loop Architecture

Creator agent produces output; Critic reviews and provides feedback; loop continues until approval or iteration limit.

criticrefinement+2
@eren9677
2025
Start Task
input
Automated Agent
agent
Human Review
human
Final Action
output

Human-in-the-Loop (HITL) Architecture

Autonomous agents collaborate with humans at critical checkpoints for approval, judgment, or handling sensitive actions.

human-in-the-loopapproval+3
@eren9677
2025
Large Search Space (e.g., Web)
input
Forager Agent 1
agent
Forager Agent 2
agent
Forager Agent 3
agent
Valuable Resource Found
tool
Aggregated Results
output

Foraging Architecture

Decentralized swarm of forager agents explore broadly; discoveries attract others to exploit rich areas.

foragingswarm-intelligence+3
@eren9677
2025
Complex Problem
input
Frontend Agent 1
agent
Frontend Agent 2
agent
Backend Agent 1
agent
Backend Agent 2
agent
Frontend Lead
agent
Backend Lead
agent

Group Architecture

Agents organized into teams with rich intra-group communication and structured inter-group coordination via liaisons.

groupteam+3
@eren9677
2025
Complex Problem
input
Orchestrator
agent
NLP Specialist Agent
agent
Computer Vision Agent
agent
Data Analyst Agent
agent
Synthesized Solution
output

Mixture of Agents

Heterogeneous agents with different specializations collaborate via an orchestrator to solve complex problems.

mixture-of-agentsheterogeneous+3
@eren9677
2025
Task Coordinator
coordinator
Agent A
worker
Agent B
worker
Agent C
worker
Agent D
worker
Result Aggregator
aggregator

Concurrent Workflows

Multiple agents work on the same task in parallel, coordinating while processing independently to reduce total time.

parallelismconcurrency+2
@eren9677
2025
Dynamic Orchestrator
orchestrator
Agent Pool
pool
Active Agent 1
active
Active Agent 2
active
Performance Monitor
monitor

Agent Rearrange

Agents dynamically add, remove, or adapt based on changing tasks and performance signals while preserving overall system integrity.

dynamicrearrange+3
@eren9677
2025
Start
start
Data Validator
processor
Security Checker
processor
Business Logic
processor
Formatter
processor
Logger
processor
End
end

Graph Workflow (DAG)

Agents are nodes and connections are edges; control flow is managed by edges and agents communicate by updating shared graph state.

graphdag+3
@eren9677
2025
Simplified view

Interactive Group Chat

An interactive group chat architecture facilitates collaboration among multiple specialized AI agents in a coordinated manner.

group-chatdynamic-speaker-selection+3
@eren9677
2025
API Endpoint
endpoint
Registry Logic
service
Agent Datastore
database
External Service
service

Agent Registry

An agent registry is a centralized architecture for managing a collection of agents. It provides functionalities to add, delete, update, and retrieve agents, ensuring that agent definitions are decoupled from their execution.

registryservice-discovery+3
@eren9677
2025
Simplified view

Spreadsheet Swarm

Spreadsheet Swarm is a multi-agent architecture designed to manage and orchestrate thousands of agents using a CSV file.

spreadsheetswarm+3
@eren9677
2025
Simplified view

Heavy Architecture

A high-performance architecture designed for handling intensive computational workloads by orchestrating multiple agents working on resource-heavy operations—perfect for large-scale data processing and high-throughput task execution.

high-performancecompute+4
@eren9677
2025
Intelligent Router
router
Task Analyzer
analyzer
Load Balancer
balancer
Routing Rules Engine
rules
NLP Agent Cluster
cluster
Vision Agent Cluster
cluster
Analytics Cluster
cluster
Workflow Cluster
cluster

Router

Priority Rules and LangGraph-style routing: analyzes tasks and routes to optimal agents/architectures based on rules, load, availability, and performance.

routerpriority-rules+3
@eren9677
2025
Simplified view

Deep Research

Specialized architecture for comprehensive, multi-domain research with iterative planning, cross-validation, and synthesis.

researchplanning+3
@eren9677
2025
Simplified view

De-Hallucination Architecture

Consensus-based validation to minimize hallucinations: primary agent generates, multiple validators fact-check, consensus decides to accept or refine.

de-hallucinationvalidation+3
@eren9677
2025
Simplified view

MALT (Multi-Agent Learning Task)

Creator produces an initial solution; Verifiers independently evaluate; Refiners improve based on feedback; iterates to high-quality output.

creatorverifier+4
@eren9677
2025
Simplified view

Majority Voting

Agents independently vote on proposed decisions; the option with the most votes is selected as final.

majority-voteconsensus+2
@eren9677
2025
Task Queue
queue
Agent 1
worker
Agent 2
worker
Agent N
worker
Round Robin Scheduler
scheduler
Processed Tasks
output

Round Robin

Tasks are distributed cyclically among agents in a fixed sequence for fair load balancing.

round-robinscheduler+2
@eren9677
2025
Task Specification
input
Builder Engine
engine
Agent Factory
factory
Configuration Store
store
Deployed Swarm
swarm

Auto-Builder

Inspects task requirements and automatically composes the necessary agents into a working swarm.

auto-buildercomposition+2
@eren9677
2025
Root Coordinator
coordinator
Peer Cluster 1
cluster
Peer Cluster 2
cluster
Agent A
agent
Agent B
agent
Agent C
agent
Agent D
agent

Hybrid Hierarchical Cluster

Top-level coordinator delegates tasks; peer clusters collaborate directly for efficiency.

hybridhierarchy+2
@eren9677
2025
Candidates
input
Agent Voter 1
voter
Agent Voter N
voter
Vote Tally
aggregator
Elected Leader
output

Election

Agents run a ballot to elect a leader or make collective choices with fault-tolerant coordination.

electionleader-election+2
@eren9677
2025
Simplified view

Dynamic Conversational

An adaptive chat architecture that selects which agents join and how messages are routed based on context, user goals, and prior dialogue—ideal for advanced customer support and context-aware chatbots.

dynamicchat+3
@eren9677
2025
Root Agent
root
Child 1
child
Child 2
child
Leaf 1
leaf
Leaf 2
leaf

Tree

A strict parent-child hierarchy where each agent has exactly one parent (except the root) and may have multiple children. This structure simplifies delegation, logging, and failure isolation.

treehierarchy+2
@eren9677
2025
User Request
input
Supervisor Agent
agent
Researcher Agent
agent
Editor Agent
agent
Writer Agent
agent
Final Output
output

Supervisor-Agent Architecture

In a Supervisor-Agent architecture, one agent acts as the central "supervisor" or "orchestrator." This supervisor agent is responsible for receiving a task, breaking it down into smaller sub-tasks, and delegating those sub-tasks to specialized "worker" agents. The supervisor then monitors the progress of the worker agents and synthesizes their outputs to produce the final result. This architecture is easy to start with and is effective for workflows that can be clearly broken down into distinct steps.

supervisororchestrator+3
@eren9677
2024