Single Agent vs Multi-Agent: When to Build a Multi-Agent System

AI Brokers

When constructing an AI agent, the design selection issues. A single agent could also be sufficient for simple duties, whereas extra advanced workflows might have a number of specialised brokers working collectively, with every one answerable for a particular a part of the method, resembling retrieval, writing, verification, coding, testing or assessment.

This put up explains the core elements of AI agent design, the ReAct method, the distinction between single-agent and multi-agent architectures, and the way to decide on the suitable design relying on the duty. It additionally features a walkthrough of how a sensible Multi-Agent RAG system works and the way it was constructed.

in style as a result of trendy LLMs at the moment are extremely succesful at duties like coding, writing, reasoning, and fixing issues throughout totally different fields. This has decreased the necessity to practice customized fashions and shifted extra consideration towards constructing sensible purposes round present LLMs. Instruments like Codex, Claude Code, Cursor and Windsurf are already serving to software program engineers work sooner, whereas companies use brokers for buyer help, automation and different real-world duties.

An AI agent is an utility that makes use of an LLM to purpose, plan and use instruments to carry out duties, permitting the mannequin to work together with its surroundings in a sensible and helpful manner.

Parts of an AI Agent

A number of the main elements of most AI brokers are the LLM, instruments, and reminiscence.

• LLM: That is the mind of the AI agent. It’s the massive language mannequin that allows the agent to purpose, plan, and determine easy methods to remedy a given process.
• Instruments: These are helpers, normally within the type of code features, that permit the LLM to work together with its surroundings. Instruments assist the agent connect with exterior information sources, search the web, retrieve data from databases, entry information, and perform particular actions. For instance, coding brokers can use instruments to jot down, debug, and save information, analysis brokers can use internet search or vector databases to assemble data and buyer help brokers can use inside firm paperwork to reply questions primarily based on trusted enterprise data.
• Reminiscence: This permits the agent to retailer related data from interactions and use it later to offer higher and extra constant help. It helps the agent preserve context throughout duties and enhance the general consumer expertise.Reminiscence could also be elective throughout early improvement, nevertheless it turns into an essential a part of many real-world AI agent methods, particularly when the agent must deal with follow-up questions, multi-step workflows or personalised interactions. There are two main kinds of reminiscence generally utilized in AI brokers: short-term reminiscence and long-term reminiscence. Quick-term reminiscence retains monitor of data inside the present session or process, whereas long-term reminiscence shops helpful data throughout a number of periods or chats so the agent can use it later.

ReAct (Reasoning + Appearing) in Brokers

An AI agent differs from a fundamental chatbot as a result of a chatbot normally follows a extra direct workflow: consumer question → LLM → response. The LLM receives the consumer’s message and generates a reply primarily based primarily on the immediate and its present context.

An AI agent goes past this by utilizing the LLM to purpose concerning the process, determine what must be accomplished, select whether or not instruments are wanted, name these instruments, observe the outcomes and proceed till it may possibly produce a helpful reply.

That is the place the ReAct method is available in. ReAct means Reasoning + Appearing. It’s an agent sample the place the LLM causes a few process and takes actions, normally by means of instruments, primarily based on that reasoning. It includes designing a core logic loop round an LLM.

A fundamental ReAct workflow in an AI agent normally seems like this:

Step 1: The agent receives a consumer question

The LLM causes over the duty and decides whether or not it may possibly reply immediately or wants to make use of instruments. It checks what instruments can be found and decides which of them are wanted to unravel the duty.

Step 2: The agent calls the required instruments

Primarily based on its reasoning, the agent takes motion by calling the mandatory instruments. These instruments could search the online, retrieve paperwork from a vector database, entry information, run code or connect with an exterior API. The outcomes returned from these instruments are often known as device outputs.

Step 3: The device outputs are despatched again to the LLM

The device outputs are handed again to the LLM as further context. This provides the agent extra related data to work with as a substitute of relying solely on the unique immediate.

Step 4: The LLM checks the proof and generates a response

The LLM evaluations the device outputs and checks whether or not they’re sufficient to unravel the duty. If the proof is enough, it generates a grounded response for the consumer. If not, the agent could repeat the reasoning, tool-calling and commentary steps till it has sufficient data to offer a helpful reply.

Construction of AI Brokers

AI Brokers can both be single or multi relying on the design construction.

Single Agent vs Multi-Agent

A single agent is an agent design the place one LLM handles the entire process. It causes, plans and calls the required instruments when wanted. Most AI brokers begin as single-agent methods as a result of they’re less complicated, simpler to keep up and normally sufficient for a lot of duties.

A multi-agent system makes use of specialised brokers to unravel totally different components of a process. It usually has a central agent, normally known as an orchestrator, supervisor or planner, that coordinates the opposite brokers and decides when every one ought to act. Every specialised agent can have its personal position, instruments and reasoning logic, making the system extra modular and appropriate for advanced workflows.

When to Construct A Multi-Agent System

A single-agent design works effectively for easy duties that require restricted device use. For instance, a private assistant agent that may entry your calendar to guide reminders, a calculator agent that solely makes use of a calculator device, or an internet search agent that makes use of an internet search API to retrieve up-to-date data.

Nonetheless, a single agent can turn into overloaded when the duty requires many instruments, multi-step reasoning, totally different tasks or verification earlier than the ultimate response is returned to the consumer. Frequent points embrace overloaded prompting, poor device routing, unclear agent tasks and decreased reliability attributable to an excessive amount of complexity in a single agent.

A multi-agent system is a more sensible choice when the duty could overwhelm a single-agent design and if you want specialised brokers with clear roles, their very own instruments and separate tasks.

For instance, a software program engineering agent may fit higher as a multi-agent system:

Orchestrator → Coder → Tester → Reviewer

The Orchestrator coordinates the workflow, the Coder agent generates the code, the Tester agent checks whether or not the code works, and the Reviewer agent evaluations the answer to verify for lacking components or potential enhancements.

One other instance is a analysis agent that researches a subject, retrieves data from totally different information sources and generates grounded content material:

Orchestrator → Retriever → Author → Verifier

The Retriever agent gathers data from the online and native paperwork saved in a vector database. The Author agent writes primarily based on the retrieved content material. The Verifier agent checks the written content material for errors, citations and factual accuracy earlier than the ultimate response is returned.

Multi-agent methods make the workflow extra modular and provides every stage a transparent position. Nonetheless, they need to be used solely when the duty genuinely wants that design, as a result of they normally enhance latency, value and upkeep complexity attributable to extra LLM calls and extra transferring components.

A easy rule is:

Use a single agent when the duty is easy, has fewer steps and desires just a few instruments. Use a multi-agent system when the duty requires specialised roles, multi-step reasoning, stronger verification or coordination throughout totally different instruments and workflows.

Walkthrough of A Multi-Agent Mission

I constructed a venture known as Multi-Agent RAG Researcher to make the thought of multi-agent methods extra sensible.

The objective of the venture is to point out how a central agent can coordinate a number of specialised brokers to analysis a subject, retrieve proof from paperwork and the online, write a grounded content material and confirm the content material earlier than returning it to the consumer. As an alternative of utilizing one agent to deal with every thing, the system splits the workflow into totally different tasks.

Check the project on github: https://github.com/ayoolaolafenwa/multi-agent-rag-researcher

Clone Mission repo

git clone https://github.com/ayoolaolafenwa/multi-agent-rag-researcher.git

Clone the repo to followup with the code alongside the put up. When the repo is cloned, the venture construction will appear like this:

.
├── docs/                         # Default PDF information
├── reminiscence/                       # SQLite-backed session reminiscence helpers
├── qdrant_vector_database/       # PDF ingestion and similarity search
├── ui/                           # Gradio app and UI handlers
├── utils/
│   ├── necessities.txt          # Python dependencies
├── worker_agents/                # Retriever, author, and verifier
├── orchestrator_agent.py         # Essential coordinator
└── run_orchestrator.py           # CLI entry level

Multi-Agent Structure

Knowledge Sources

There are two main information sources:

1. Qdrant Vector Database

Info retrieval from PDFs is dealt with within the following levels:

• A number of PDFs could be loaded from the docs/ folder or uploaded by means of the UI.
• Paperwork are cut up into chunks, transformed into embeddings, and saved in an area Qdrant assortment.
• Similarity search is then used to retrieve probably the most related chunks throughout the listed paperwork.
• The retrieved chunks embrace quotation metadata resembling doc identify and web page quantity.

The doc retrieval a part of the venture the place Qdrant vector database is setup, PDF ingestion, chunking, embedding, and similarity search are managed is dealt with in qdrant_vector_database/vector_store.py .

2. Tavily Internet Search

Tavily is used to retrieve up-to-date or exterior data from the online. The retriever agent can use it when:

• the listed PDFs don’t cowl the question
• doc proof is weak or incomplete
• newer data is required

Employee Brokers

1. Retriever Agent

The position is:

• It makes use of two instruments: PDF doc retrieval and internet search.
• Given a question, it decides whether or not to make use of native paperwork, internet search or each.
• If native doc proof is lacking or weak, it may possibly fall again to internet search to assemble broader or extra up-to-date context.

The code for the retriever agent with tavily internet search out there in worker_agents/retriever.py . It makes use of gpt-5.4-mini with low reasoning effort.

2. Author Agent

The position is:

• It receives the retrieved data from the Retriever Agent.
• It writes a grounded draft primarily based on the out there proof.
• It contains supporting citations from PDFs or internet sources when they’re out there.

The code for the author agent out there in worker_agents/writer.py . It makes use of gpt-5.4 with low reasoning effort.

3. Verifier Agent

The position is:

• It receives the draft from the Author Agent along with the proof.
• It checks whether or not the claims within the draft are supported by the retrieved proof.
• It returns the ultimate verified response.

The code for the employee agent is on the market in worker_agents/verifier.py . It makes use of gpt-5.4 with low reasoning effort.

Reminiscence

SQLite is used to offer short-term reminiscence for the multi-agent workflow. For a given session ID, the system shops:

• the newest consumer question
• the newest retrieved proof for that session

This permits the orchestrator to reuse related proof for follow-up questions as a substitute of retrieving the identical data once more each time.

The code for the reminiscence is on the market in memory/memory.py .

Orchestrator

The orchestrator coordinates the three employee brokers: Retriever, Author and Verifier.

How the Orchestrator coordinates the Multi-Agent Workflow

• It receives the consumer question and, relying on the question, could reply immediately or start the evidence-based workflow.
• For a analysis question, it first checks whether or not related cached proof from the reminiscence for the present session could be reused.
• If cached proof just isn’t sufficient, it calls the Retriever Agent to assemble proof from PDFs, the online or each.
• If there’s doc proof however the proof is weak, the Retriever Agent may also fetch up-to-date data from the online to complement the native doc data.
• The orchestrator then passes the lively proof and the consumer question to the Author Agent so it may possibly generate a grounded draft.
• Subsequent, it sends the draft and proof to the Verifier Agent, which checks the claims and returns the ultimate verified report.
• In the course of the session, the newest question and retrieved proof are saved in reminiscence for follow-up questions.
• In follow-up questions, the orchestrator could reuse cached proof as a substitute of calling the Retriever Agent once more, then proceed with the Author Agent and Verifier Agent to generate the ultimate response.

The code for the orchestrator is in orchestrator_agent.py . It makes use of gpt-5.4-mini with low reasoning effort.

The orchestrator has a guardrail that retains the system centered on analysis and factual questions. It refuses unrelated normal duties resembling coding assist or basic math as a result of the objective of the system is to perform as a analysis assistant.

Observe: For the fashions used within the orchestrator and employee brokers, you’ll be able to change them from gpt-5.4 to any openai offered mannequin of your selection.

Mission Setup

Stipulations

• Python 3.10 or newer
• OpenAI API key: Create an OpenAI Account should you don’t have one and Generate an API Key.
• Tavily API key: Tavily is a specialised web-search device for AI brokers. Create an account on Tavily.com, as soon as your profile is ready up, an API key will likely be generated you could copy into your surroundings. New account receives 1000 free credit that can be utilized for as much as 1000 internet searches.

Set up

1. Create and activate a digital surroundings:

python3 -m venv env
supply env/bin/activate

2. Set up the dependencies:

cd multi-agent-rag-researcher
pip3 set up -r utils/necessities.txt

3. Create a utils/var.env file and retailer your API keys:

OPENAI_API_KEY=your_openai_api_key
TAVILY_API_KEY=your_tavily_api_key

4. Place the PDFs you wish to index within the docs/ folder, or add PDFs later by means of the UI. The venture already contains present PDFs in docs/, at the moment Gemma 3 Technical Report.pdf and DeepSeek-V3.2.pdf, so you should utilize these immediately or change them with your individual paperwork.

Run Mission

Begin the command-line app:

python3 run_orchestrator.py

When the CLI begins, it ingests the PDFs in docs/ into the native Qdrant retailer. Sort q or exit to finish the session.

Run UI for Multi-Agent Chat

Begin the Gradio UI:

python3 ui/gradio_app.py

The UI routinely masses the default PDFs from docs/ on startup. If you happen to add new PDFs, they change the lively listed doc set for that UI session.

Demo Video of the Multi Agent Agent RAG Researcher

Notes

• Session reminiscence is saved in utils/reminiscence.db.
• Native Qdrant information is saved in utils/qdrant_storage/.
• The system is designed for analysis and factual query answering, not for unrelated general-purpose duties.

Conclusion

On this put up, I defined how an AI agent works, the way it makes use of instruments to work together with its surroundings, and the way the ReAct method helps it purpose, plan, choose instruments and execute particular duties.

I additionally lined the structural design of AI brokers, which could be single-agent or multi-agent methods. I defined how each designs work, when to decide on every one primarily based on the workflow, and in contrast single-agent implementation with multi-agent structure.

Lastly, I did a walkthrough of the multi-agent design behind my Multi-Agent RAG Researcher venture, exhibiting the way it makes use of an orchestrator to coordinate three employee brokers, retrieve data from the online and native paperwork, use reminiscence for consistency and write and confirm grounded content material earlier than returning the ultimate output.

Attain to me through:

E mail: [email protected]

Linkedin: https://www.linkedin.com/in/ayoola-olafenwa-003b901a9/

References

https://developers.openai.com/cookbook

https://developers.openai.com/api/docs/guides/function-calling