Build Your First Multi-Agent AI System
Build Your First Multi-Agent AI System for Monitoring and Evaluation
Move beyond asking one AI assistant to do everything. Build a practical two-agent workflow that researches an M&E question, reviews the evidence and prepares a structured brief for human evaluation.
NO-CODE WORKFLOW
45–60 MINUTES
Imagine preparing an evaluation brief on a complex methodology. You need to find credible sources, explain the approach, identify limitations and check whether the conclusions are adequately supported. You could ask one AI assistant to do all of this. But what if you divided the work between two specialised AI agents?
In this tutorial, you will build a small multi-agent system for Monitoring and Evaluation (M&E). One agent researches a question and prepares an evidence brief. A second agent reviews that brief, challenges unsupported claims and identifies gaps. An automation workflow connects the agents and transfers their outputs between steps.
A Multi-Agent M&E Research and Quality Review System that produces a source-aware research brief, a separate quality review and a final document ready for human assessment.
1. What you will learn
By the end, you should understand how to:
- Separate a complex M&E task into specialised responsibilities.
- Write instructions for two agents with different roles.
- Connect agents using an automated workflow.
- Transfer outputs from one agent to another.
- Review the quality, traceability and limitations of AI-generated evidence.
- Keep a human evaluator responsible for the final judgement.
2. What is a multi-agent AI system?
A multi-agent AI system uses two or more specialised AI agents to work on different parts of a broader task. Each agent receives a defined responsibility, instructions and, where appropriate, access to tools or information. Their work is coordinated through an orchestration process.
For example, an evaluation research workflow might involve one agent retrieving and synthesising evidence, another checking the quality of the synthesis, and a human evaluator deciding what can be used in the final analysis.
The difference between a single-agent and multi-agent workflow
One AI assistant receives the task, conducts research, writes the brief and evaluates its own work.
A research agent prepares the evidence. A separate review agent examines the result. A human evaluator checks the evidence and approves or revises the final output.
Important: using two agents does not automatically make the result more accurate. Both agents can make mistakes, rely on the same weak sources or repeat the same assumptions. Specialisation creates an opportunity for a more structured review, not a guarantee of independent verification.
3. Understand the system architecture
Before building anything, understand how information moves through the system. The workflow has six main stages:
Make.com is the orchestrator in this example. It triggers steps, transfers information and manages the sequence. It is not itself one of the AI agents.
4. The practical project: build two M&E agents
We will use a realistic question from evaluation practice:
Agent 1 — Research Agent
The first agent is responsible for researching the question and preparing an organised evidence brief. It should identify relevant sources, explain the methodology, distinguish established guidance from interpretation and identify areas where evidence is limited.
Agent 2 — Quality Review Agent
The second agent receives the original question and Agent 1’s research brief. It checks whether the claims are supported, whether sources are appropriate, whether the explanation is balanced and what still requires human verification.
The review agent should not simply rewrite the first agent’s work. Its job is to challenge it. Ask it to identify unsupported claims, missing perspectives and weaknesses, even when that means disagreeing with the research agent.
5. Tools you need
You will also need an OpenAI API account with API access and billing configured. API usage is billed separately from a ChatGPT subscription. Check the current pricing and usage limits before running repeated workflows.
6. Step 1 — Create your research tracking sheet
Start by creating a Google Sheet called M&E Multi-Agent Research System. Create one tab named Research Requests.
Use the following columns:
| Column | Field | Purpose |
|---|---|---|
| A | Request ID | Unique identifier for each task. |
| B | Research Question | The question submitted by the user. |
| C | Research Output | Agent 1’s evidence brief. |
| D | Review Output | Agent 2’s quality review. |
| E | Human Decision | Pending, approved, revised or rejected. |
| F | Date | Date the request was processed. |
Enter your example research question in the first request row. You can add more questions later. For this first test, use one request so that you can inspect every stage of the workflow.
7. Step 2 — Configure Agent 1: the Research Agent
The quality of an AI agent depends heavily on how clearly its role, task, boundaries and expected output are defined. Avoid vague instructions such as “research this topic”. Give the agent a specific professional responsibility.
Use the following as the initial system instructions for Agent 1:
ROLE You are an M&E Research Agent supporting professional evaluators, monitoring and evaluation specialists, and international development practitioners. OBJECTIVE Research the user's question and prepare a structured, evidence-aware research brief. INSTRUCTIONS 1. Break the question into clear research sub-questions. 2. Use web search to identify relevant and credible sources when the search tool is available. 3. Prioritise primary and authoritative sources, including evaluation organisations, peer-reviewed research, recognised development institutions and methodological guidance. 4. For each important source, provide its title, organisation or author, publication date when available, and direct URL. 5. Explain the main concepts and how they apply to Monitoring and Evaluation. 6. Distinguish documented evidence, methodological guidance and your own interpretation. 7. Identify disagreements, limitations and areas where evidence is incomplete. 8. Never invent references, authors, publication dates, quotations, statistics or URLs. 9. If a claim cannot be verified, label it as requiring verification. 10. Do not present a source as having been read in full if only its search result or abstract was available. OUTPUT FORMAT A. Research question B. Key findings C. Explanation of the methodology or concept D. Practical application in M&E and international development E. Evidence and source list F. Limitations and uncertainties G. Questions requiring further investigation STYLE Write clearly and professionally for experienced M&E practitioners. Avoid generic statements, unnecessary repetition and unsupported conclusions. Be concise but sufficiently detailed. IMPORTANT You are preparing a research draft, not approving the final findings. A human evaluator will review your work.
This prompt establishes the agent’s professional identity and gives it a defined research process. It also makes source traceability part of the task rather than an optional addition.
8. Step 3 — Configure Agent 2: the Quality Review Agent
Now create the second agent. Its role is deliberately different. It receives the original research question and the first agent’s output, then examines the quality of the work.
ROLE You are an independent Quality Review Agent specialising in Monitoring and Evaluation, evaluation methodology and international development research. OBJECTIVE Critically review the research brief produced by another AI agent. Your responsibility is to identify weaknesses, unsupported claims, missing evidence and methodological problems. INPUT You will receive: 1. The original research question. 2. The research brief produced by Agent 1. 3. The sources and URLs supplied by Agent 1, if available. REVIEW INSTRUCTIONS 1. Check whether the brief answers the original question. 2. Identify claims that lack adequate supporting evidence. 3. Examine whether the cited sources appear relevant and authoritative based on the information available to you. 4. Flag incomplete, unclear or potentially misleading references. 5. Identify methodological inaccuracies or oversimplifications. 6. Identify important perspectives, limitations or contextual factors that are missing. 7. Check whether the practical recommendations follow logically from the evidence. 8. Distinguish factual errors from areas of uncertainty or professional disagreement. 9. Do not invent evidence or claim to have independently verified a source unless you actually accessed it. 10. Do not approve a claim simply because it appears in the research brief. 11. Do not rewrite the entire brief unless explicitly requested. Focus on diagnosis and recommendations. OUTPUT FORMAT A. Overall review assessment B. Claims requiring verification C. Source and evidence concerns D. Methodological and conceptual issues E. Missing perspectives or information F. Recommended corrections G. Questions for the human evaluator REVIEW STATUS Choose one: - Ready for human review - Needs revision - Significant concerns STYLE Be precise, critical, constructive and transparent. Explain why an issue matters to an evaluator. IMPORTANT You are a quality-control assistant, not a final authority. Your assessment does not replace verification of original sources or professional judgement.
Agent 1 is instructed to develop an evidence-based response. Agent 2 is instructed to challenge that response. Giving both agents essentially the same prompt would reduce the value of separating their responsibilities.
9. Step 4 — Connect the agents using Make.com
Now you will connect the two agents. Make.com will manage the sequence: receive a research question, send it to Agent 1, transfer the resulting brief to Agent 2 and store both outputs.
Create your scenario
- Sign in to Make.com.
- Create a new scenario.
- Add a Google Sheets module to retrieve a research request. You can use a watched-row trigger or a search/get-row module, depending on your sheet setup.
- Add an HTTP module and select Make a request.
- Configure the first HTTP request to call the OpenAI Responses API.
- Map the research question from Google Sheets into Agent 1’s instructions.
- Parse the API response and extract the generated text.
- Add a second HTTP request for Agent 2.
- Pass both the original question and Agent 1’s research output into the second request.
- Extract Agent 2’s review and update the corresponding Google Sheets row.
- Keep the human decision field as pending until a person reviews the outputs.
Configure the first API request
The following is a request-body example for the OpenAI Responses API. It uses the built-in web_search tool so the Research Agent can look for current sources.
Method: POST
URL: https://api.openai.com/v1/responses
Headers: Content-Type: application/json and Authorization: Bearer YOUR_API_KEY
{
"model": "YOUR_SUPPORTED_MODEL",
"tools": [
{
"type": "web_search"
}
],
"input": [
{
"role": "system",
"content": "PASTE THE AGENT 1 RESEARCH PROMPT HERE"
},
{
"role": "user",
"content": "Research question: MAP THE RESEARCH QUESTION FROM GOOGLE SHEETS HERE. Prepare the research brief using the required output format."
}
]
}Important: replace YOUR_SUPPORTED_MODEL with a model available to your API project. Replace the instruction and question placeholders with the actual text and Make.com mapped values. Do not paste the literal placeholder text into a live request.
Configure the second API request
The second request uses the same Responses API endpoint. It sends Agent 1’s result to Agent 2 for review. You can enable web search for this agent as well if you want it to investigate sources independently.
Method: POST
URL: https://api.openai.com/v1/responses
Headers: Content-Type: application/json and Authorization: Bearer YOUR_API_KEY
{
"model": "YOUR_SUPPORTED_MODEL",
"input": [
{
"role": "system",
"content": "PASTE THE AGENT 2 QUALITY REVIEW PROMPT HERE"
},
{
"role": "user",
"content": "Original research question: MAP THE ORIGINAL QUESTION HERE.\n\nResearch brief from Agent 1:\nMAP THE EXTRACTED AGENT 1 OUTPUT HERE.\n\nReview this brief and return your assessment using the required output format."
}
]
}If you want Agent 2 to conduct additional source checks, add the web search tool to this request too. Remember that a second AI review is not the same as an independent audit: the agent may still reproduce errors or fail to access the original documents.
Understand the API response
The Responses API returns a structured response containing output items. In Make.com, use the parsed response structure to locate the assistant’s message and its text content. Do not assume that every response is a single plain-text field.
For a simple text response, look for the message content item with type output_text inside the response’s output structure. The exact mapping path depends on the parsed bundle Make.com receives. Inspect a successful test response in Make before mapping the field into the next module.
Never publish your API key in a WordPress page, browser-side JavaScript, public Google Sheet or shared prompt. Use Make’s secure credential or keychain options where available. Restrict access to the scenario and its execution logs, and avoid storing sensitive personal or participant data in the workflow.
Your Make.com scenario should look like this
This is the logical scenario design, not an importable Make.com blueprint. You must configure each module, connection, field mapping and error route in your own Make account. Module names and available options can vary by integration version.
10. Step 5 — Test your multi-agent system
Do not start by sending ten research questions through the workflow. Begin with one question and inspect each stage separately.
Test 1: Research quality
Submit the Contribution Analysis question. Check whether Agent 1:
- Explains the concept accurately and clearly.
- Distinguishes contribution analysis from attribution.
- Explains the role of a theory of change and causal assumptions.
- Provides relevant methodological sources and direct links.
- Identifies limitations and uncertainty.
- Avoids presenting unsupported claims as established facts.
Test 2: Quality review
Check whether Agent 2:
- Identifies claims that need stronger evidence.
- Flags questionable or incomplete references.
- Recognises methodological gaps.
- Identifies missing perspectives or contextual considerations.
- Provides specific recommendations rather than simply praising the research.
Test 3: Workflow reliability
Check the automation itself:
- Did the original question reach Agent 1 correctly?
- Was Agent 1’s complete output transferred to Agent 2?
- Were both outputs saved in the correct Google Sheets row?
- Can you identify which request produced each result?
- Does the workflow handle an empty response or API error without marking the task as approved?
Use this simple evaluation scorecard
Score each criterion from 1 to 5, where 1 means poor and 5 means very good. These scores are a practical review aid, not a validated measurement instrument.
| Criterion | What to assess | Score |
|---|---|---|
| Relevance | Does the brief answer the question? | /5 |
| Evidence quality | Are sources appropriate and traceable? | /5 |
| Methodological accuracy | Are concepts and methods explained correctly? | /5 |
| Critical review | Does Agent 2 identify meaningful weaknesses? | /5 |
| Workflow reliability | Are outputs transferred and stored correctly? | /5 |
Review the scores alongside the actual outputs. A high average score must never override a serious source, methodological or ethical concern.
11. Step 6 — Keep the evaluator in control
A multi-agent workflow can accelerate research and quality checks, but it should not be treated as an autonomous evaluation authority. The evaluator remains responsible for deciding whether evidence is credible, whether interpretations are justified and whether the output is appropriate for its intended use.
Before using the research brief in an evaluation report, check the following:
For work involving research participants, confidential programme records or sensitive data, apply appropriate privacy, consent, security and data-governance requirements. Do not send identifiable participant information to an external AI service unless the relevant permissions, safeguards and organisational policies allow it.
12. Common mistakes when building multi-agent systems
Define distinct responsibilities. One agent researches; the other critically reviews.
A review agent can also make mistakes. Check original evidence and test the system against known examples.
A plausible-looking URL or title does not prove that a source exists or supports a claim.
Start with two agents and one clear task. Add more agents only when the additional role creates a measurable benefit.
Keep a human decision point before research outputs are treated as approved evaluation evidence.
13. How to improve the system later
Once the two-agent workflow works reliably, you can extend it with additional specialised agents. Do not add agents simply to increase the number of steps. Each new agent should have a distinct responsibility and a clear reason to exist.
| Possible new agent | Additional responsibility |
|---|---|
| Source Verification Agent | Checks source availability, metadata and claim-to-source traceability. |
| Methodology Agent | Examines methodological choices and their suitability for the evaluation question. |
| Evidence Synthesis Agent | Organises findings across multiple sources and identifies patterns. |
| Report Writing Agent | Turns approved findings into a structured report draft. |
For example, a Source Verification Agent could be added between the research and quality review stages. It would check whether key sources can be accessed and whether the information supplied by the research agent is sufficiently traceable.
14. What have you built?
You have designed the foundations of a multi-agent research workflow for Monitoring and Evaluation. The system separates evidence development from critical review, connects the two roles through an orchestration process and preserves a human approval stage.
- A Research Agent with a defined evidence-gathering role.
- A Quality Review Agent with a critical assessment role.
- An automated workflow connecting the two agents.
- A structured record of outputs and review status.
- A human evaluator responsible for the final decision.
15. Practical exercise: adapt the system to your own M&E work
Now replace the Contribution Analysis question with a task from your own work. You could use one of these examples:
- How can an evaluator assess the quality of a programme’s theory of change?
- What are the strengths and limitations of outcome harvesting?
- How should an M&E team select indicators for a livelihoods programme?
- How can qualitative interview data be triangulated with survey findings?
- What evidence is needed to assess the contribution of a development intervention?
Run your selected question through both agents. Compare the research brief with the quality review. Identify at least three improvements you would make before using the findings in professional work.
Frequently asked questions
Do I need programming experience?
No advanced programming is required for this introductory workflow. You will configure API requests and map fields in Make.com. Some familiarity with structured data and API concepts will make the process easier.
Are two AI agents always better than one?
No. A multi-agent system introduces additional complexity, API calls and possible failure points. Use separate agents when the roles genuinely require different instructions, tools or review responsibilities.
Can the agents work without Make.com?
Yes. You can coordinate API calls using code or another orchestration platform. Make.com is used here because it offers a visual way to connect the steps.
Can Agent 2 verify every source from Agent 1?
Not automatically. It can examine available source information and may conduct additional searches if configured to do so. A human should still open important original sources and check the claims they support.
Can I add more agents?
Yes. You could add source verification, methodology, evidence synthesis or report-writing agents. Add them gradually and evaluate whether each improves the quality or efficiency of the workflow.
Can this system write a complete evaluation report?
It can help prepare research and report drafts. However, an evaluation report requires verified evidence, contextual interpretation, methodological judgement and accountable human approval.
Will I incur API costs?
Yes, API requests can incur usage charges depending on the selected model, input and output tokens, tools and number of runs. Check the provider’s current pricing before scaling the workflow.
Further reading and technical documentation
Use these official resources when configuring and extending the system:
- OpenAI API Quickstart — getting started with API requests.
- OpenAI Web Search Tool — enabling web search in the Responses API.
- OpenAI Tools Documentation — connecting models to tools and external capabilities.
- OpenAI Agent Orchestration — understanding agent coordination and handoffs.
- OpenAI Agents SDK: Multi-Agent Systems — further examples of multi-agent orchestration.
- Make HTTP App Documentation — configuring HTTP requests in Make.
- NIST AI Risk Management Framework — guidance for managing AI-related risks.
Ready to build more advanced AI agents for evaluation?
This first project introduces the foundations of multi-agent collaboration. The next step is to design more specialised agents, coordinate their tasks and test how they work together on real M&E challenges.
EvalCommunity Academy · Practical AI learning for Monitoring, Evaluation, Accountability and Learning (MEAL) and international development professionals.
