MS Foundry: Workflow failing to invoke agent when the agent is created with the Python SDK: "The requested operation requires an element of type 'Object', but the target element has type 'Array'."

I hit this exact error when I started creating agents through the Python SDK.

The root cause is simple: the workflow engine expects the agent definition to have an input_schema (an Object), but agents created through the SDK default to an empty Array instead.

Agents created in the Portal automatically generate the correct schema, so they work.

When the workflow tries to invoke your agent, it reads the agent’s metadata, sees an Array where it expects an Object, and throws:

“The requested operation requires an element of type ‘Object’, but the target element has type ‘Array’.”

The fix is to explicitly define the input schema when creating the agent. If you don’t, the SDK produces an empty list for input_schema, which breaks workflow invocation.

Here’s the minimal fix:

from azure.ai.projects.models import PromptAgentDefinition, JSONSchema

agent = project_client.agents.create_version(
    agent_name = agent_name,
    definition = PromptAgentDefinition(
        model = agent_model,
        instructions = instructions,
        tools = [mcp_kb_tool],
        input_schema = JSONSchema(
            type="object",
            properties={
                "input": JSONSchema(type="string")
            },
            required=["input"]
        )
    )
)

After adding the schema, the workflow will invoke the agent without that error.

That was the only change I needed when I ran into this.

Thanks for the reply! MS seems to have changed the schema for the latest version I'm using of the SDK (azure-ai-projects=2.0.0b2), so can't use input_schema or JSONSchema since they don't exist anymore. I gave this a try which I think is equivalent to what you did. It executes successfully, but I still got the same error. I wish MS docs had an example about how to use this properly.

agent = project_client.agents.create_version(
    agent_name = agent_name,
    definition = PromptAgentDefinition(
        model = agent_model,
        instructions = instructions,
        tools = [mcp_kb_tool],
        structured_inputs = {
            "input": StructuredInputDefinition(
                description = "User input text",
                required = True
            )
        }
    )
)

Hi Jason Lee,

Thank you for reaching out, and I understand how frustrating this error can be when working with workflows!

The error you're seeing indicates that the weekend-planner step is returning an Array (list of results), but the End step expects a single Object (JSON object). This is a workflow configuration issue, not a problem with agent creation.

Here are three ways to resolve this:

Modify the workflow output mapping: In the workflow designer, configure the weekend-planner step to extract the first item from the array before passing it to the End step

Add a transformation step: Insert a step between weekend-planner and End that converts the array to a single object

Update the End step: Configure the End step to accept an array instead of an object

Regarding your agent creation code:

Your agent creation approach looks correct. Here's a complete working example for reference:

from azure.ai.projects import AIProjectClient
from azure.ai.projects.models import MCPTool, PromptAgentDefinition
from azure.identity import DefaultAzureCredential

# ========== CONFIGURATION ==========
PROJECT_ENDPOINT = "<YOUR-PROJECT-ENDPOINT>"

PROJECT_CONNECTION_NAME = "<KNOWLEDGE_BASE_CONNECTION_NAME>"  # The connection name you created

AGENT_NAME = "<AGENT-NAME-TO-CREATE>"
AGENT_MODEL = "gpt-4.1" 


MCP_ENDPOINT = "<KNOWLDGE-BASE-URL>"

INSTRUCTIONS = """
You are an AI assistant that answers questions using the connected knowledge base.
Provide accurate answers based only on the retrieved content.
"""

# ===================================

def main():
    credential = DefaultAzureCredential()

    project_client = AIProjectClient(
        endpoint=PROJECT_ENDPOINT,
        credential=credential
    )

    # Create MCP Tool linked to Knowledge Base
    mcp_kb_tool = MCPTool(
        server_label="knowledge-base",
        server_url=MCP_ENDPOINT,
        require_approval="never",
        allowed_tools=["knowledge_base_retrieve"],
        project_connection_id=PROJECT_CONNECTION_NAME
    )

    # Create Agent
    agent = project_client.agents.create_version(
        agent_name=AGENT_NAME,
        definition=PromptAgentDefinition(
            model=AGENT_MODEL,
            instructions=INSTRUCTIONS,
            tools=[mcp_kb_tool],
        ),
    )

    print("Agent created successfully!")
    print("Agent ID:", agent.id)


if __name__ == "__main__":
    main()

To fix your specific workflow issue, could you please share:The code where agent is invoked (you can send via private message if preferred):

This will help me provide more specific guidance on resolving the array-to-object conversion issue.

Thank you for reaching out to The Microsoft Q&A Portal.

Hi Aryan, thanks for the reply. Sure, here's the YAML from the workflow invoking this agent. I removed the end conversation step to eliminate that as a factor, but the problem still occurs. I also looked at the YAML of a portal-created workflow and added some things that were missing from my work flow (e.g. output messages: Local.LatestMessage) but that didn't solve the problem either.

kind: workflow
trigger:
  kind: OnConversationStart
  id: trigger_wf
  actions:
    - kind: InvokeAzureAgent
      id: Plan-Weekend
      agent:
        name: weekend-planner
      description: Looks at event listings for the coming weekend to answer the user's question
      conversationId: =System.ConversationId
      input:
        messages: =System.LastMessage
      output:
        messages: Local.LatestMessage
        autoSend: true
id: ""
name: Weekend-planner-flow
description: ""

Hi Jason Lee,
I understand this process can be a bit challenging, and I appreciate your efforts so far. When you get a chance, could you please try building the agent using the Python SDK, as demonstrated in the code I shared above