How to pass outputs from multiple agents as input to a downstream agent in Azure AI Foundry workflows

Hi S K, Raghav

In Connected Agents, the intended design is: you have a primary (main) agent that delegates to multiple connected sub‑agents, and then the main agent compiles the sub‑agents’ responses into a single final response. That means the “merge” happens inside the main agent orchestration, not by you concatenating outputs externally. The doc explicitly notes that connected agent responses are only visible to the main agent, and the main agent “will compile the responses from the connected agents and provide the response.” [learn.microsoft.com]

So, if your goal is: Agent‑3 needs both outputs → Make Agent‑3 be the “main agent”, and connect Agent‑1 and Agent‑2 as its connected agents. Then Agent‑3 can ask each one for its part and synthesize.

# multiple_connected_agents.py

from azure.identity import DefaultAzureCredential

from azure.ai.agents import AgentsClient  # adjust import to your SDK preview

# Hypothetical helper types; align with your SDK build:

from azure.ai.agents.tools import ConnectedAgentTool  # adjust if your SDK uses a different namespace

 

cred = DefaultAzureCredential()

client = AgentsClient(endpoint="<YOUR_PROJECT_ENDPOINT>", credential=cred)

 

# Assume these two connected agents already exist (created earlier):

stock_price_agent = client.administration.get_agent(agent_name="StockPriceAgent")

company_news_agent = client.administration.get_agent(agent_name="CompanyNewsAgent")

 

2. How to “merge” Agent‑1 + Agent‑2 for Agent‑3 using Connected Agents (recommended)**

• Create Agent‑1 (specialist)
• Create Agent‑2 (specialist)
• Create Agent‑3 (main orchestrator) and add Agent‑1 and Agent‑2 in Connected agents section (portal) or via ConnectedAgentTool (Python). [learn.microsoft.com]
• In Agent‑3 instructions, tell it explicitly to:
  1. call Agent‑1
     1. call Agent‑2
        1. combine both results and answer in a structured way.

Because in connected agents, the main agent naturally routes tasks using natural language and composes results—no custom orchestration logic required. [learn.microsoft.com]

3) Python example: Agent‑3 as main agent + two connected agents

This is the core idea from the doc’s Python pattern, extended to two sub‑agents. (The doc shows one connected agent; the same approach applies to multiple sibling sub‑agents.) [learn.microsoft.com]

 

Python

 

import os
from azure.ai.projects import AIProjectClient
from azure.ai.agents.models import ConnectedAgentTool, MessageRole
from azure.identity import DefaultAzureCredential
project_client = AIProjectClient(
    endpoint=os.environ["PROJECT_ENDPOINT"],
    credential=DefaultAzureCredential(),
)

# Agent 1 (specialist)

agent1 = project_client.agents.create_agent(

    model=os.environ["MODEL_DEPLOYMENT_NAME"],

    name="agent_1",

    instructions="You produce Output A. Return ONLY Output A in a concise form."

)

# Agent 2 (specialist)

agent2 = project_client.agents.create_agent(

    model=os.environ["MODEL_DEPLOYMENT_NAME"],

    name="agent_2",

    instructions="You produce Output B. Return ONLY Output B in a concise form."

)

# Connect both agents as tools

tool1 = ConnectedAgentTool(id=agent1.id, name=agent1.name, description="Produces Output A")

tool2 = ConnectedAgentTool(id=agent2.id, name=agent2.name, description="Produces Output B")

# Agent 3 (main orchestrator)

agent3 = project_client.agents.create_agent(

    model=os.environ["MODEL_DEPLOYMENT_NAME"],

    name="agent_3_orchestrator",

    instructions=(

        "You are Agent 3. You MUST:\n"

        "1) Call agent_1 to get Output A.\n"

        "2) Call agent_2 to get Output B.\n"

        "3) Merge A and B into a single combined analysis.\n"

        "Return the final answer with sections: OutputA, OutputB, Combined."

    ),

    tools=[*tool1.definitions, *tool2.definitions],

)

thread = project_client.agents.threads.create()

project_client.agents.messages.create(

    thread_id=thread.id,

    role=MessageRole.USER,

    content="Run the workflow. I need both outputs combined for downstream decision."

)

run = project_client.agents.runs.create_and_process(thread_id=thread.id, agent_id=agent3.id)

print(run.status)

Show more lines

Key point: you’re not passing two separate “inputs” into Agent‑3. Instead, Agent‑3 asks both connected agents and merges internally, as designed. [learn.microsoft.com]

If you are in a Workflow/Logic App situation where “InvokeAzureAgent” allows only one messages input

You mentioned “merging Agent‑1 and Agent‑2 input for Agent input”. That exact pain appears often in workflow orchestration: Agent‑3 accepts only one messages variable, so you can’t map two variables directly.

Supported workaround (practical + safe)

Create a single text message by aggregating both outputs with clear delimiters, then send that as the one messages payload.

Use a structure like:

Plain Text

SYSTEM: You are Agent 3. Use OutputA and OutputB below.
USER:
OutputA:
<<<
{agent1_output}
>>>

OutputB:
<<<
{agent2_output}
>>>

Task:
Merge OutputA and OutputB and produce the final result.

This is the simplest “merge” that stays compatible with “single messages input” constraints (and avoids passing arrays / complex structures that often fail in strict message normalization environments). A very similar scenario is described in community discussions where the action only accepts one messages input and the fix is an aggregation step. [learn.microsoft.com]

5) Best-practice prompt for Agent‑3 to merge without losing separation

When you concatenate, don’t just append strings. Use a contract:

• Tell Agent‑3: “treat OutputA and OutputB as separate sources”
• Ask it to produce:
  • “What came from A”
  • “What came from B”
  • “What is combined / resolved”
  • “Conflicts / assumptions”

Important limitations from the Connected Agents doc

A few constraints matter when designing merges:

• Depth limit = 2: main agent can have multiple connected agents (siblings), but sub‑agents can’t have their own sub‑agents. [learn.microsoft.com]
• Citations may not pass through reliably from connected agents to the main agent. [learn.microsoft.com]
• Connected agents can’t call local functions via function calling tool; use OpenAPI/Azure Functions instead. [learn.microsoft.com]

Please let me know and if there are any remaining questions or additional details, I can help with, I’ll be glad to provide further clarification or guidance.

Thankyou!

Hi S K, Raghav

Did you get any chance to review the above response

Thank you!

I will check and get back to you on this

Thanks for the guidance.

To align with the single messages input constraint, I have updated the workflow to aggregate outputs from multiple upstream agents using a SetVariable step that produces a single plain-text message with clear delimiters. This aggregated text is then passed as the sole messages input to the downstream agent.

At this stage, the implementation is done entirely using Azure AI Foundry workflow YAML (SetVariable + InvokeAzureAgent), and I am not using the Python SDK.

Could you please confirm whether this aggregation pattern using SetVariable and a single plain-text messages payload is the recommended and supported approach for multi-agent workflows in Teams/Copilot scenarios?

Hi S K, Raghav

I have gone through your use case. As per my understanding, you have two different agents in place having different set of deliverables and you want to pass the outcomes from those two agents to third agent.

Most feasible way was to configure through SDK.

The Hands-off approach mentioned in this blog looks more plausible to your use case.

https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/building-a-multimodal-multi-agent-system-using-azure-ai-agent-service-and-openai/4396267

I am trying to get more context on Sequential Agent workflow to check the feasibility of mapping the approach of concatenating the results from both Agents and have requested your availability on teams call for the same in private message.

Requesting your inputs on the same.