Multi Agent Systems: Applications & Comparison of Tools

Multi-agent systems(MAS) enable distinct AI agents to work together to achieve complex objectives. Every AI agent in the system possesses its specific characteristics and responsibilities that contribute to a greater goal. MAS provides a distinctive approach to managing multi-step tasks and enhancing efficiency..

We examined the most popular multi-agent system platforms and 5 case studies for businesses of all types that want to implement multi-agent systems in their workflows.

Comparison of conversational & orchestration platforms

We have established a benchmark for a travel planning multi-agent system. The agents in our system utilize OpenAI’s o4-mini model for web searching, while the booking agents operate with a random simulator that has a 70% failure rate, requiring the agent to make multiple attempts to complete the booking. All of the platforms below passed these tests:

• Integration with a free search tool (DuckDuckGo search)
• Finding various options based on requests, such as a 2-day trip with a maximum 4-hour flight from Paris within a budget of €500, and a 5-day trip with a maximum 2-hour flight from San Francisco with in a budget of $1000.
• Sorting by different criteria like flight length, price, or hotel ratings, and creating reservations based on those criteria.

The complexity of implementing vendors is assessed based on these criteria:

• What amount of code is needed?
• How straightforward is the agent definition process?
• How effortless is it to debug and track execution?

Key features of the vendors are presented as follows:

LangChain

Architecture & Orchestration: LangChain features a chain-based architecture linking components sequentially, supporting simple chains and complex agent systems.

Tool Integration: LangChain integrates external tools and APIs via its “Tools” framework.

Memory & State Management: The platform offers different memory systems (buffer, conversation, entity) for chains.

Best For:

• Designing intricate AI workflows
• Applications that necessitate integration with various external tools
• Projects that demand flexibility in selecting LLM providers

Choose LangChain if:

• You require adaptable solutions and comprehensive tool integration.
• Your project demands a robust ecosystem.
• You wish to rapidly prototype using various LLMs.

Google Agent2Agent

Architecture & Orchestration: Agent2Agent focuses on enabling communication between autonomous agents. Unlike LangChain’s sequential approach, it emphasizes message-passing between peer agents.

Tool Integration: Offers tool integration, but to a lesser extent. Typically requires custom development.

Memory & State Management: Provides basic memory capabilities, often requiring custom implementation for complex state tracking.

Best For:

• Research applications that examine collaboration among agents
• Situations where the independence of agents takes precedence over prescribed workflows

Choose Agent2Agent if:

• Your research focuses on the cooperation of agents.
• You are investigating emergent behaviors.
• You require a peer-to-peer agent architecture.

CrewAI

Architecture & Orchestration: CrewAI presents the “crew” model where agents assume defined roles and responsibilities. This role-focused strategy is highly effective for intricate workflows that demand specialized expertise.

Tool Integration: Strong tool integration capabilities that focus on tools appropriate for each role.

Memory & State Management: Establishes memory systems based on each agent’s specific roles.

Best For:

• Complex workflows that demand various specialized roles
• Initiatives that emulate human organizational frameworks
• Processes that gain from distinct delineation of responsibilities

Choose CrewAI if:

• Your workflow mirrors human organizational frameworks.
• You necessitate a clear separation of agent roles.
• Delegating complex tasks is central to your application.

Microsoft Bot SDK

Architecture & Orchestration: Focuses on the idea of bots and dialogues, emphasizing a more organized method for managing conversation flow. While it is less adaptable to intricate multi-agent systems, it excels in established conversational frameworks.

Tool Integration: High compatibility with the Microsoft ecosystem (Azure services, Office 365, etc.), but less versatile for general-purpose tools.

Memory & State Management: Offers strong state management via its Bot State service, enabling continuity throughout conversation turns.

Best For:

• Business applications that demand close integration with Microsoft
• Support bots designed for consistent dialogue flows
• Projects that utilize the Azure ecosystem

Choose Microsoft Bot SDK if:

• You’re already invested in the Microsoft ecosystem
• Enterprise integration is a priority
• You are developing conversational bots for customer interaction.

Rasa

Architecture & Orchestration: Rasa adopts an NLU-first strategy combined with a policy-driven system for action determination. This approach relies on a more conventional conversational AI framework, distinguishing it from the more recent LLM-based methodologies.

Tool Integration: Limited native tool integration, while custom actions enable connections to external systems.

Memory & State Management: Utilizes tracker stores to preserve conversation state and context throughout dialogue sessions.

Best For:

• On-premise deployments with privacy requirements
• Applications that demand precise control over natural language understanding (NLU)
• Projects that need to integrate rule-based and machine learning (ML) methods

Choose Rasa if:

• On-premise deployment is necessary.
• You require detailed control over NLU.
• Privacy and data sovereignty are essential.

Other multi-agent systems platforms

Enterprise integration & automation

• IBM Cloud Pak for Integration (Watson Orchestrator) leverages containerized agents, each encapsulating connectors, rules engines, or low‑code flows that can be chained together in an end‑to‑end automation pipeline.
• MuleSoft Anypoint Platform offers “autonomous integration agents” that run API‑led connectivity tasks and can coordinate via a central Anypoint Exchange registry.

Robotics & IoT

• AWS RoboMaker provides simulation and fleet management for multiple ROS‑based robot agents that share maps, tasks, and sensor data via AWS IoT Core.
• Microsoft Azure Digital Twins models physical environments as collections of interacting “twins” (agents) that can invoke each other’s APIs for telemetry, control logic, and event routing.

Decentralized & research‐oriented frameworks

• JADE (Java Agent DEvelopment Framework) is a FIPA‑compliant, open‑source Java stack from Telecom Italia Labs for building and running heterogeneous agent applications.
• SingularityNET is a blockchain-based marketplace where autonomous AI services publish capabilities, negotiate tasks, and settle transactions in AGI tokens.

Key concepts of multi-agent systems

1. Multi‑agent reinforcement learning (MARL) enables each agent to learn the best behaviors via trial and error, adapting its policies based on the actions of other agents and changes in the environment. By exchanging observations or gradients, agents collaboratively develop strategies, occasionally displaying emergent phenomena termed “autocurricula,” in which new skills emerge through competition or cooperation among agents.¹
2. Agents operate in a shared environment, using broadcasting or direct messaging to coordinate on common tasks. These communication methods, spanning from blackboard systems to publish-subscribe brokers, enable agents to collaborate without needing a central controller.²
3. Independent agents, despite having different goals, can still collaborate by negotiating task allocations (e.g., through the Contract Net Protocol). This allows them to dynamically assign roles and resources, addressing challenges that exceed the capabilities of any one agent.³
4. Interactions among multiple agents act as robust models for complex systems, simulating everything from financial markets to social network dynamics and disaster‑response logistics. They offer valuable insights into the macroscopic behaviors that arise from straightforward agent rules.

What are multi agent systems?

A multi-agent system is a computerized system in which multiple agents interact with each other to achieve common or conflicting goals. Multi-agent systems research focuses on designing and developing systems that can solve complex problems by coordinating the actions of multiple agents.

Developing multi-agent systems necessitates a thorough knowledge of artificial intelligence, distributed systems, and agent systems. Autonomous agents excel at tackling intricate tasks that demand collaboration and coordination between several agents.

Multi-agent systems can effectively model and simulate real-world scenarios, such as supply chain and transportation management. Decentralized agents can collaboratively pursue shared objectives, even in dynamic settings with restricted communication.

While single agents prolong the process by creating excess information and working one by one instead of triggering a “chain reaction. ” Also, an ecosystem of single agents can create unstable outcomes for the next process, whereas multi-agent systems control the outcomes to be suitable for each other.

Well-designed multi-agent systems consist of proactive autonomous agents, meaning problems seldom arise and the execution of the desired process is disrupted minimally, if at all.

System architecture of multi agent systems

Hierarchical structure

Hierarchical structure is one of the most popular methods. In this method, supervisor agents assign subtasks to subordinate workers, facilitating scalable task division and management.⁴

In this framework, agents typically become specialists; one might focus on web scraping, while another oversees data analysis, each customized for its specific domain and deployed as microservices or encapsulated modules.⁵

Peer‑to‑Peer (Flat) Structure

In P2P MAS, all agents function equally, with no central authority. They dynamically find peers through methods like gossip or lightweight registries, and negotiate tasks directly, such as using the Contract Net. This decentralized structure excels in dynamic environments like ad‑hoc IoT networks, drone swarms, or peer‑to‑peer energy markets, allowing nodes to join, leave, and share services independently. To reduce coordination overhead, designers frequently incorporate local caching, adaptive polling, or lightweight aggregation layers.⁶

LLM‑Mediated Hub

Large language models (LLMs) have allowed agents to utilize transformer‑based reasoning elements to enhance decision‑making processes and facilitate natural language interactions, serving as “universal translators” that connect diverse agents and human participants in hybrid environments and ecosystems.⁷

Core components & architecture patterns

• Agent kernels integrate planning, reasoning, and execution components, frequently utilizing a belief‑desire‑intention (BDI) engine or policy network to govern agent actions.
• Message buses and brokers (such as pub/sub systems, message queues, or service meshes) offer reliable, decoupled channels for event propagation and request management.
• Registry and discovery services facilitate dynamic agent identification and authentication through standards like DNS‑SD, mDNS, or proprietary registries, ensuring agents can securely find and interact with peers in real time.

Communication & coordination mechanisms

Multi‑agent systems utilize various interaction models: request-response for synchronous queries, publish–subscribe for event streams, and blackboard systems for collaborative shared state, each customized to distinct latency and coupling needs. 

Classic coordination patterns consist of the Contract Net Protocol (where managers solicit proposals, contractors bid, and contracts are granted), auction mechanisms for allocating resources, and leader election algorithms for dynamically assigning control roles among peers.

Agents share structured data using schemas and serialization formats, such as JSON‑LD for linked data interoperability, Protocol Buffers for efficient binary messaging, and RDF for semantic web integration.

Case studies of multi-agent systems

Invoice processing of Matrix

A pipeline of agents, an OCR/ data‑extraction agent, a business‑rule validation agent, an approval‑routing agent, and a payment‑execution agent, operates collaboratively under a central orchestrator. Each agent makes decisions, like flagging anomalies or selecting approval paths, and publishes events on a message bus; downstream agents subscribe and react, creating a loosely coupled workflow.

Matrix, a multi‑agent LLM framework that progressively develops domain expertise through memory modules, has demonstrated significant improvements in handling complex invoice fields in collaboration with a global logistics company.⁸

Smart buildings of OpenBlue

In Johnson Controls OpenBlue, individual agents handle HVAC, lighting, occupancy sensing, and security. A central “supervisor” agent oversees key performance indicators (KPIs) such as energy consumption and comfort, dynamically adjusting set‑points for each subsystem agent.⁹

Instead of being a single entity, every building function operates as a separate service‑agent, trading resources through a publish–subscribe system to achieve safety and efficiency objectives.

Warehouse robot swarm of Amazon

Amazon’s Kiva “Drive Units” operate as a decentralized fleet: navigation agents exchange map fragments, collision‑avoidance agents communicate local intents, and task‑assignment agents auction shelf‑retrieval jobs.¹⁰

There isn’t a singular controller; instead, robots interact in real time to prevent traffic jams and enhance throughput, demonstrating emergent flow patterns that boost efficiency.

Economic policy simulation of CANVAS

The Bank of Canada’s CANVAS framework models households, firms, banks, and a central bank agent. Firms and consumers independently decide on spending and investments; banks determine lending rates; the central bank agent adjusts its policy rule.¹¹

Each sector‑agent adheres to its unique behavioral rules, exchanging price and quantity messages at every “tick,” allowing macro outcomes to arise from micro interactions.

Misinformation containment on X

In an X simulation, agents termed “debunkers” identify likely false posts, while “amplifiers” disseminate corrections, and “monitors” measure the rate of spread. A coordination agent assigns debunker resources to the most viral threads.¹²

Each type of agent operates under its specific behavioral rules, exchanging messages about price and quantity at each “tick,” allowing macro outcomes to arise from micro interactions.

Benefits, challenges & limitations of multi‑agent systems

Benefits

• Scalability: Horizontally adding specialized agents allows for meeting increasing demands without creating central bottlenecks. 
• Resilience: Isolating faults and enabling dynamic fail‑over guarantees that failures of individual agents won’t bring down the entire system. 
• Modularity: Loosely coupled and language‑agnostic components support incremental upgrades and diverse deployments. 

Challenges & limitations

• Complexity: Emergent behaviors and non-deterministic systems make debugging and validation more challenging.
• Standards fragmentation: Competing protocols such as FIPA ACL, KQML, and A2A may impede adoption and interoperability.
• Security & privacy: Robust authentication and encryption measures are required to build trust among autonomous agents, counter threats from malicious actors, and protect sensitive data.

Best Practices for Designing MAS

Establish precise agent contracts through standardized communication languages (like FIPA ACL) and thoroughly documented API schemas or SLAs, ensuring clarity regarding each agent’s inputs, outputs, and potential failure modes.

Utilize a service mesh (such as Istio or Linkerd) to oversee inter‑agent communications, providing observability, traffic control, and load balancing while keeping agent code complexity minimal.

Introduce health checks and circuit breakers at the infrastructure layer to swiftly identify unresponsive agents and prevent cascading failures. This will allow the system to gracefully degrade under heavy load.

Further readings

• Agentic Document Extraction

External Links

• 1. https://arxiv.org/pdf/2312.10256
• 2. Multi-agent system - Wikipedia. Contributors to Wikimedia projects
• 3. https://web.media.mit.edu/~cynthiab/Readings/tan-MAS-reinfLearn.pdf
• 4. What are hierarchical multi-agent systems?.
• 5. What is a Multiagent System? | IBM.
• 6. SmythOS - Multi-Agent System Architecture: Building Blocks for Effective Collaboration. SmythOS, LLC
• 7. [2503.13514] RAG-KG-IL: A Multi-Agent Hybrid Framework for Reducing Hallucinations and Enhancing LLM Reasoning through RAG and Incremental Knowledge Graph Learning Integration.
• 8. [2412.15274] Memory-Augmented Agent Training for Business Document Understanding.
• 9. OpenBlue Digital Twin Overview | Johnson Controls.
• 10. Amazon steps up use of robotics in warehouses. Financial Times
• 11. https://www.bankofcanada.ca/wp-content/uploads/2022/12/swp2022-51.pdf
• 12. [2502.01450] Simulating Rumor Spreading in Social Networks using LLM Agents.