A Full Information to Constructing Multi-Agent Methods

Trendy AI functions depend on clever brokers that assume, cooperate, and execute complicated workflows, whereas single-agent methods wrestle with scalability, coordination, and long-term context. AgentScope AI addresses this by providing a modular, extensible framework for constructing structured multi-agent methods, enabling function task, reminiscence management, instrument integration, and environment friendly communication with out pointless complexity for builders and researchers alike looking for sensible steerage at present now clearly. On this article, we offer a sensible overview of its structure, options, comparisons, and real-world use instances.

What’s AgentScope and Who Created It?

AgentScope is an open-source multi-agent framework for AI agent methods that are structured, scalable, and production-ready. Its predominant focus is on clear abstractions, modular design together with communication between brokers relatively than ad-hoc immediate chaining. 

The AI methods group’s researchers and engineers primarily created AgentScope to beat the obstacles of coordination and observability in intricate agent workflows. The truth that it may be utilized in analysis and manufacturing environments makes it a rigour-laden, reproducible and extensible framework that may nonetheless be dependable and experimental on the similar time. 

Additionally Learn: Single-Agent vs Multi-Agent Methods

Why AgentScope Exists: The Downside It Solves

As LLM functions develop extra complicated, builders more and more depend on a number of brokers working collectively. Nonetheless, many groups wrestle with managing agent interactions, shared state, and long-term reminiscence reliably. 

AgentScope solves these issues by introducing express agent abstractions, message-passing mechanisms, and structured reminiscence administration. Its core objectives embrace: 

• Transparency and Flexibility: The whole functioning of an agent’s pipeline, which incorporates prompts, reminiscence contents, API calls, and power utilization, is seen to the developer. You’re allowed to cease an agent in the course of its reasoning course of, test or change its immediate, and proceed execution with none difficulties. 
• Multi-Agent Collaboration: Relating to performing difficult duties, the necessity for a number of specialised brokers is most well-liked over only one huge agent. AgentScope has built-in assist for coordinating many brokers collectively. 
• Integration and Extensibility: AgentScope was designed with extensibility and interoperability in thoughts. It makes use of the newest requirements just like the MCP and A2A for communication, which not solely permit it to attach with exterior providers but additionally to function inside different agent frameworks. 
• Manufacturing Readiness: The traits of many early agent frameworks didn’t embrace the potential for manufacturing deployment. AgentScope aspires to be “production-ready” proper from the beginning. 

In conclusion, AgentScope is designed to make the event of complicated, agent-based AI methods simpler. It gives modular constructing blocks and orchestration instruments, thus occupying the center floor between easy LLM utilities and scalable multi-agent platforms. 

Core Ideas and Structure of AgentScope

• Agent Abstraction and Message Passing: AgentScope symbolizes each agent as a standalone entity with a particular operate, psychological state, and choice-making course of. Brokers don’t change implicit secret context, thus minimizing the incidence of unpredictable actions. 
• Fashions, Reminiscence, and Instruments: AgentScope divides intelligence, reminiscence, and execution into separate elements. This partitioning permits the builders to make modifications to every half with out disrupting your complete system. 
• Mannequin Abstraction and LLM Suppliers: AgentScope abstracts LLMs behind a consolidated interface, henceforth permitting easy transitions between suppliers. Builders can select between OpenAI, Anthropic, open-source fashions, or native inference engines. 
• Brief-Time period and Lengthy-Time period Reminiscence: AgentScope differentiates between short-term conversational reminiscence and long-term persistent reminiscence. Brief-term reminiscence gives the context for instant reasoning, whereas long-term reminiscence retains data that lasts. 
• Software and Operate Invocation: AgentScope offers brokers the chance to name exterior instruments through structured operate execution. These instruments may encompass APIs, databases, code execution environments, or enterprise methods. 

Key Capabilities of AgentScope

AgentScope is an all-in-one bundle of a number of highly effective options which permits multi-agent workflows. Listed below are some principal strengths of the framework already talked about:  

• Multi-Agent Orchestration: AgentScope is a grasp within the orchestration of quite a few brokers working to realize both overlapping or opposing objectives. Furthermore, the builders have the choice to create a hierarchical, peer-to-peer, or perhaps a coordinator-worker method.  

async with MsgHub(
    contributors=[agent1, agent2, agent3],
    announcement=Msg("Host", "Introduce yourselves.", "assistant"),
) as hub:
    await sequential_pipeline([agent1, agent2, agent3])

    # Add or take away brokers on the fly
    hub.add(agent4)
    hub.delete(agent3)

    await hub.broadcast(Msg("Host", "Wrap up."), to=[])

• Software Calling and Exterior Integrations: AgentScope has a easy and simple integration with the exterior methods through instrument calling mechanisms. This characteristic helps to show brokers from easy conversational entities into environment friendly automation elements that perform actions.  
• Reminiscence Administration and Context Persistence: With AgentScope, the builders have the ability of explicitly controlling the context of the brokers’ storage and retrieval. Thus, they determine what info will get retained and what will get to be transient. The advantages of this transparency embrace the prevention of context bloating, fewer hallucinations, and reliability in the long run. 

QuickStart with AgentScope

For those who observe the official quickstart, the method of getting AgentScope up and operating is sort of easy. The framework necessitates Python model 3.10 or above. Set up will be carried out both via PyPI or from the supply:

From PyPI:

Run the next instructions within the command-line:

pip set up agentscope 

to put in the latest model of AgentScope and its dependencies. (If you’re utilizing the uv atmosphere, execute uv pip set up agentscope as described within the docs) 

From Supply:  

Step 1: Clone the GitHub repository: 

git clone -b predominant https://github.com/agentscope-ai/agentscope.git 
cd agentscope 

Step 2: Set up in editable mode: 

pip set up -e . 

This can set up AgentScope in your Python atmosphere, linking to your native copy. You can even use uv pip set up -e . if utilizing an uv atmosphere.  

After the set up, you must have entry to the AgentScope lessons inside Python code. The Hiya AgentScope instance of the repository presents a really fundamental dialog loop with a ReActAgent and a UserAgent.  

AgentScope doesn’t require any further server configurations; it merely is a Python library. Following the set up, it is possible for you to to create brokers, design pipelines, and do some testing instantly. 

Making a Multi-Agent Workflow with AgentScope

Let’s create a practical multi-agent system by which two AI fashions, Claude and ChatGPT, possess completely different roles and compete with one another: Claude generates issues whereas GPT makes an attempt to unravel them. We will clarify every a part of the code and see how AgentScope really manages to carry out this interplay. 

1. Setting Up the Atmosphere 

Importing Required Libraries 

import os
import asyncio
from typing import Checklist

from pydantic import BaseModel
from agentscope.agent import ReActAgent
from agentscope.formatter import OpenAIChatFormatter, AnthropicChatFormatter
from agentscope.message import Msg
from agentscope.mannequin import OpenAIChatModel, AnthropicChatModel
from agentscope.pipeline import MsgHub

All the mandatory modules from AgentScope and Python’s commonplace library are imported. The ReActAgent class is used to create the clever brokers whereas the formatters be certain that messages are ready accordingly for the varied AI fashions. Msg is the communication methodology between brokers offered by AgentScope. 

Configuring API Keys and Mannequin Names 

os.environ["OPENAI_API_KEY"] = "your_openai_api_key"
os.environ["ANTHROPIC_API_KEY"] = "your_claude_api_key"

OPENAI_API_KEY = os.environ["OPENAI_API_KEY"]
ANTHROPIC_API_KEY = os.environ["ANTHROPIC_API_KEY"]

CLAUDE_MODEL_NAME = "claude-sonnet-4-20250514"
GPT_SOLVER_MODEL_NAME = "gpt-4.1-mini"

This setup will assist in authenticating the API credentials for each OpenAI and Anthropic. And to entry a specific mannequin we’ve to cross the precise mannequin’s title additionally.  

2. Defining Information Constructions for Monitoring Outcomes 

Spherical Log Construction: 

class RoundLog(BaseModel):
    round_index: int
    creator_model: str
    solver_model: str
    downside: str
    solver_answer: str
    judge_decision: str
    solver_score: int
    creator_score: int

This knowledge mannequin holds all the knowledge concerning each spherical of the competition in real-time. Collaborating fashions, generated issues, solver’s suggestions, and present scores are being recorded thus making it straightforward to evaluate and analyze every interplay. 

International Rating Construction: 

class GlobalScore(BaseModel):
    total_rounds: int
    creator_model: str
    solver_model: str
    creator_score: int
    solver_score: int
    rounds: Checklist[RoundLog]

The general competitors outcomes throughout all rounds are stored on this construction. It preserves the ultimate scores and your complete rounds historical past thus providing us a complete view of brokers’ efficiency within the full workflow. 

Normalizing Agent Messages 

def extract_text(msg) -> str:
    """Normalize an AgentScope message (or comparable) right into a plain string."""
    if isinstance(msg, str):
        return msg

    get_tc = getattr(msg, "get_text_content", None)
    if callable(get_tc):
        textual content = get_tc()
        if isinstance(textual content, str):
            return textual content

    content material = getattr(msg, "content material", None)
    if isinstance(content material, str):
        return content material

    if isinstance(content material, checklist):
        components = []
        for block in content material:
            if isinstance(block, dict) and "textual content" in block:
                components.append(block["text"])
        if components:
            return "n".be part of(components)

    text_attr = getattr(msg, "textual content", None)
    if isinstance(text_attr, str):
        return text_attr

    messages_attr = getattr(msg, "messages", None)
    if isinstance(messages_attr, checklist) and messages_attr:
        final = messages_attr[-1]
        last_content = getattr(final, "content material", None)
        if isinstance(last_content, str):
            return last_content

        last_text = getattr(final, "textual content", None)
        if isinstance(last_text, str):
            return last_text

    return ""

Our operate here’s a supporting one that enables us to acquire readable textual content from agent responses with reliability whatever the message format. Completely different AI fashions have completely different buildings for his or her responses so this operate takes care of all of the completely different codecs and turns them into easy strings we are able to work with. 

4. Constructing the Agent Creators 

Creating the Downside Creator Agent (Claude) 

def create_creator_agent() -> ReActAgent:
    return ReActAgent(
        title="ClaudeCreator",
        sys_prompt=(
            "You're Claude Sonnet, performing as an issue creator. "
            "Your job: in every spherical, create ONE practical on a regular basis downside that "
            "some individuals may face (e.g., scheduling, budgeting, productiveness, "
            "communication, private resolution making). "
            "The issue ought to:n"
            "- Be clearly described in 3–6 sentences.n"
            "- Be self-contained and solvable with reasoning and customary sense.n"
            "- NOT require personal knowledge or exterior instruments.n"
            "Return ONLY the issue description, no answer."
        ),
        mannequin=AnthropicChatModel(
            model_name=CLAUDE_MODEL_NAME,
            api_key=ANTHROPIC_API_KEY,
            stream=False,
        ),
        formatter=AnthropicChatFormatter(),
    )

This utility produces an assistant that takes on the function of Claude and invents practical issues of on a regular basis life that aren’t essentially such. The system immediate specifies the type of issues to be created, primarily making it the situations the place reasoning is required however no exterior instruments or personal info are required for fixing them. 

Creating the Downside Solver Agent (GPT) 

def create_solver_agent() -> ReActAgent:
    return ReActAgent(
        title="GPTSolver",
        sys_prompt=(
            "You're GPT-4.1 mini, performing as an issue solver. "
            "You'll obtain a sensible on a regular basis downside. "
            "Your job:n"
            "- Perceive the issue.n"
            "- Suggest a transparent, actionable answer.n"
            "- Clarify your reasoning in 3–8 sentences.n"
            "If the issue is unclear or inconceivable to unravel with the given "
            "info, you MUST explicitly say: "
            ""I can't remedy this downside with the knowledge offered.""
        ),
        mannequin=OpenAIChatModel(
            model_name=GPT_SOLVER_MODEL_NAME,
            api_key=OPENAI_API_KEY,
            stream=False,
        ),
        formatter=OpenAIChatFormatter(),
    )

This instrument additionally offers delivery to a different agent powered by GPT-4.1 mini whose predominant job is to discover a answer to the issue. The system immediate dictates that it should give a transparent answer together with the reasoning, and most significantly, to acknowledge when an issue can’t be solved; this frank recognition is crucial for correct scoring within the competitors. 

5. Implementing the Judging Logic 

Figuring out Answer Success 

def solver_succeeded(solver_answer: str) -> bool:
    """Heuristic: did the solver handle to unravel the issue?"""
    textual content = solver_answer.decrease()

    failure_markers = [
        "i cannot solve this problem",
        "i can't solve this problem",
        "cannot solve with the information provided",
        "not enough information",
        "insufficient information",
    ]

    return not any(marker in textual content for marker in failure_markers)

This judging operate is easy but highly effective. If the solver has really offered an answer or confessed failure the operate will test. By looking for sure expressions that present the solver was not capable of handle the difficulty, the winner of each spherical will be determined routinely with out the necessity for human intervention. 

6. Operating the Multi-Spherical Competitors 

Most important Competitors Loop 

async def run_competition(num_rounds: int = 5) -> GlobalScore:
    creator_agent = create_creator_agent()
    solver_agent = create_solver_agent()
    creator_score = 0
    solver_score = 0
    round_logs: Checklist[RoundLog] = []

    for i in vary(1, num_rounds + 1):
        print(f"n========== ROUND {i} ==========n")

        # Step 1: Claude creates an issue
        creator_msg = await creator_agent(
            Msg(
                function="consumer",
                content material="Create one practical on a regular basis downside now.",
                title="consumer",
            ),
        )

        problem_text = extract_text(creator_msg)
        print("Downside created by Claude:n")
        print(problem_text)
        print("n---n")

        # Step 2: GPT-4.1 mini tries to unravel it
        solver_msg = await solver_agent(
            Msg(
                function="consumer",
                content material=(
                    "Right here is the issue it's essential to remedy:nn"
                    f"{problem_text}nn"
                    "Present your answer and reasoning."
                ),
                title="consumer",
            ),
        )

        solver_text = extract_text(solver_msg)
        print("GPT-4.1 mini's answer:n")
        print(solver_text)
        print("n---n")

        # Step 3: Decide the consequence
        if solver_succeeded(solver_text):
            solver_score += 1
            judge_decision = "Solver (GPT-4.1 mini) efficiently solved the issue."
        else:
            creator_score += 1
            judge_decision = (
                "Creator (Claude Sonnet) will get the purpose; solver failed or admitted failure."
            )

        print("Decide resolution:", judge_decision)
        print(f"Present rating -> Claude: {creator_score}, GPT-4.1 mini: {solver_score}")

        round_logs.append(
            RoundLog(
                round_index=i,
                creator_model=CLAUDE_MODEL_NAME,
                solver_model=GPT_SOLVER_MODEL_NAME,
                downside=problem_text,
                solver_answer=solver_text,
                judge_decision=judge_decision,
                solver_score=solver_score,
                creator_score=creator_score,
            )
        )

    global_score = GlobalScore(
        total_rounds=num_rounds,
        creator_model=CLAUDE_MODEL_NAME,
        solver_model=GPT_SOLVER_MODEL_NAME,
        creator_score=creator_score,
        solver_score=solver_score,
        rounds=round_logs,
    )

    # Last abstract print
    print("n========== FINAL RESULT ==========n")
    print(f"Whole rounds: {num_rounds}")
    print(f"Creator (Claude Sonnet) rating: {creator_score}")
    print(f"Solver (GPT-4.1 mini) rating: {solver_score}")

    if solver_score > creator_score:
        print("nOverall winner: GPT-4.1 mini (solver)")
    elif creator_score > solver_score:
        print("nOverall winner: Claude Sonnet (creator)")
    else:
        print("nOverall consequence: Draw")

    return global_score

This represents the core of our multi-agent course of. Each spherical Claude proposes a difficulty, GPT tries to unravel it, and we determine the scores are up to date and every little thing is logged. The async/await sample makes the execution easy, and after all of the rounds are over, we current the whole outcomes that point out which AI mannequin was general higher. 

7. Beginning the Competitors 

global_result = await run_competition(num_rounds=5)

This single assertion is the start line of your complete multi-agent competitors for five rounds. Since we’re utilizing await, this runs completely in Jupyter notebooks or different async-enabled environments, and the global_result variable will retailer all of the detailed statistics and logs from your complete competitors 

Actual-World Use Circumstances of AgentScope 

AgentScope is a extremely versatile instrument that finds sensible functions in a variety of areas together with analysis, automation, and company markets. It may be deployed for each experimental and manufacturing functions. 

• Analysis and Evaluation Brokers: The very first space of utility is analysis evaluation brokers. AgentScope is among the finest options to create a analysis assistant agent that may accumulate info with none assist.  
• Information Processing and Automation Pipelines: One other attainable utility of AgentScope is within the space of knowledge processing and automation. It could handle pipelines the place the info goes via completely different levels of AI processing. In this type of system, one agent may clear knowledge or apply filters, one other may run an evaluation or create a visible illustration, and a 3rd one may generate a abstract report. 
• Enterprise and Manufacturing AI Workflows: Lastly, AgentScope is created for high-end enterprise and manufacturing AI functions. It caters to the necessities of the true world via its options which might be built-in: 
  • Observability 
  • Scalability 
  • Security and Testing 
  • Lengthy-term Initiatives 

When to Select AgentScope 

AgentScope is your go-to answer once you require a multi-agent system that’s scalable, maintainable, and production-ready. It’s a sensible choice for groups that have to have a transparent understanding and oversight. It might be heavier than the light-weight frameworks however it’s going to undoubtedly repay the hassle when the methods turn out to be extra difficult. 

• Undertaking Complexity: In case your utility actually requires the cooperation of a number of brokers, such because the case in a buyer assist system with specialised bots, or a analysis evaluation pipeline, then AgentScope’s built-in orchestration and reminiscence will assist you a large number. 
• Manufacturing Wants: AgentScope places an amazing emphasis on being production-ready. For those who want robust logging, Kubernetes deployment, and analysis, then AgentScope is the one to decide on.  
• Expertise Preferences: In case you’re utilizing Alibaba Cloud or want assist for fashions like DashScope, then AgentScope can be your good match because it gives native integrations. Furthermore, it’s suitable with commonest LLMs (OpenAI, Anthropic, and so forth.).  
• Management vs Simplicity: AgentScope offers very detailed management and visibility. If you wish to undergo each immediate and message, then it’s a really appropriate selection. 

Extra Examples to Strive On

Builders take the chance to experiment with concrete examples to get essentially the most out of AgentScope and get an perception into its design philosophy. Such patterns characterize typical cases of agentic behaviors. 

• Analysis Assistant Agent: The analysis assistant agent is able to find sources, condensing the outcomes, and suggesting insights. Assistant brokers confirm sources or present counter arguments to the conclusions. 
• Software-Utilizing Autonomous Agent: The autonomous tool-using agent is ready to entry APIs, execute scripts and modify databases. A supervisory agent retains observe of the actions and checks the outcomes. 
• Multi-Agent Planner or Debate System: The brokers working as planners give you methods whereas the brokers concerned within the debate problem the assumptions. A choose agent amalgamates the ultimate verdicts. 

Conclusion

AgentScope AI is the proper instrument for making scalable and multi-agent methods which might be clear and have management. It’s the finest answer in case a number of AI brokers have to carry out the duty collectively, with no confusion in workflows and mastery of reminiscence administration. It’s using express abstractions, structured messaging, and modular reminiscence design that brings this know-how ahead and solves lots of points which might be generally related to prompt-centric frameworks. 

By following this information; you now have a whole comprehension of the structure, set up, and capabilities of AgentScope. For groups constructing large-scale agentic functions, AgentScope acts as a future-proof method that mixes flexibility and engineering self-discipline in fairly a balanced manner. That’s how the multi-agent methods would be the predominant a part of AI workflows, and frameworks like AgentScope would be the ones to set the usual for the subsequent technology of clever methods. 

Often Requested Questions