Share via

Azure AI Foundry - Workflow feature - Parallel agents and upload files

Jean Lanza 60 Reputation points
2026-03-17T13:55:56.02+00:00

I wann know if there is an option to upload files to be processed by a workflow and also an option to use agents in parallel on a workflow, for example:
User's image

Foundry Agent Service
Foundry Agent Service

A fully managed platform in Microsoft Foundry for hosting, scaling, and securing AI agents built with any supported framework or model

Answer accepted by question author
  1. Q&A Assist
    2026-03-17T13:56:13.0533333+00:00
    AI generated content. This question contains an answer created with AI from Q&A Assist. Learn more

    Workflows in Azure AI Foundry are designed exactly for orchestrating multiple agents and other steps in a predefined process. They support:

    1. Multi‑agent orchestration in a workflow
      Workflow-oriented multi-agent patterns are explicitly supported. A workflow engine such as Microsoft Foundry workflows can orchestrate multiple agents as separate steps in a process. Each step can call a different agent, with explicit sequencing and guards (preconditions, post-conditions, thresholds). The architecture guidance also notes that multi-agent workflows can be executed serially or in parallel, using sequential or concurrent orchestration patterns. This matches the pattern in the diagram where Agent 1 and Agent 2 run as separate steps and then converge into a later step (such as “Format”).
    2. Parallel agent execution pattern
      Workflow-oriented multi-agent solutions can be run concurrently when appropriate. The guidance describes this as concurrent orchestration, where multiple agent steps are invoked in parallel and their results are combined later in the workflow. This is suitable for scenarios like incident triage, ETL, or multi-stage processing where different agents can work independently on the same or related inputs before a downstream step aggregates or formats the results.
    3. Uploading and using files with agents in workflows
      Azure AI Foundry agents can use the file search tool backed by vector stores:
      • Files are uploaded via the Files API.
      • A vector store is created and associated with those file IDs.
      • An agent is created with the file search tool configured to use that vector store.
      • The agent then answers questions or performs tasks using the uploaded file content.
      In a workflow, a typical pattern is:
      • Early step: upload files and create/update a vector store.
      • Subsequent agent steps: call agents that are configured with that vector store (or thread-level vector stores) so they can retrieve and reason over the uploaded content.
    4. File upload and cleanup operations
      The file search documentation shows the concrete REST operations used to:
      • Upload files (/openai/v1/files).
      • Create vector stores and associate file IDs (/openai/v1/vector_stores).
      • Use those vector stores from agents with the file search tool.
        It also shows how to delete agents, vector stores, and files when the workflow is complete.

    Putting this together, a workflow like the one in the diagram is supported conceptually:

    • Start → step that uploads files and prepares a vector store.
    • Parallel steps → Agent 1 and Agent 2, both configured to use the same or different vector stores, potentially executed concurrently.
    • Merge step → a “Format” step that aggregates outputs from both agents and produces the final result.

    References:

    0 comments

2 additional answers

Sort by: Most helpful
Most helpful Newest Oldest
  1. Karnam Venkata Rajeswari 1,310 Reputation points Microsoft External Staff Moderator
    2026-03-17T14:11:02.4066667+00:00

    Hello Jean Lanza,

    Welcome to Microsoft Q&A .Thank you for reaching out.

    Based on the provided workflow diagram, this represents a valid Azure AI Foundry workflow pattern where multiple agents are orchestrated as part of a single process.

    Azure AI Foundry workflows are designed to orchestrate multiple agents either sequentially or in parallel. Parallel execution is supported through concurrent orchestration, where more than one agent step can run at the same time and complete independently. Once all parallel steps finish, the workflow can continue to a downstream step that aggregates or transforms the outputs. This aligns directly with the diagram where Agent 1 and Agent 2 run in parallel and then converge into a formatting step.

    File inputs are not processed directly by the workflow engine itself. Instead, files are uploaded using the file upload and vector store mechanism, and agents are configured to access those files through the file search tool. A typical design involves uploading files first, creating or updating a vector store, and then invoking one or more agents in the workflow that can retrieve and reason over that content. This allows the same uploaded content to be shared across multiple agents running in parallel.

    A recommended high‑level approach is:

    • Upload and store file inputs using the file upload and vector store pattern

    • Configure agents with access to the same vector store

    • Invoke multiple agents in parallel within the workflow

    • Merge and format agent outputs in a final step for a single response

    This approach aligns with supported multi‑agent orchestration patterns and provides a clear, scalable way to process shared inputs and deliver a single, well‑structured outcome through Azure AI Foundry workflows.

    References:

    How to use Azure OpenAI Assistants file search (classic) - Microsoft Foundry (classic) | Microsoft Learn

    Introducing Multi-Agent Workflows in Foundry Agent Service | Microsoft Foundry Blog

    Build a workflow in Microsoft Foundry - Microsoft Foundry | Microsoft Learn

    Thank you!

    1 person found this answer helpful.
    0 comments
  2. Manas Mohanty 16,035 Reputation points Microsoft External Staff Moderator
    2026-03-25T17:43:18.2+00:00

    Hi Jean Lanza,

    Below are supported Multi Agent workflowUser's image Most of above workflow suitable except Sequential workflow (Agents works in Sequence)

    Concurrent workflow

    Concurrent Orchestration

    Group Chat workflow

    Group Chat Context Synchronization

    Handoff workflow

    Handoff Orchestration

    https://dotnet.territoriali.olinfo.it/en-us/agent-framework/workflows/orchestrations/

    You can use same file search tool in all Agents in context.

    # Create vector store and upload file
vector_store = openai.vector_stores.create(name="ProductInfoStore")

with asset_file_path.open("rb") as file_handle:
    vector_store_file = openai.vector_stores.files.upload_and_poll(
        vector_store_id=vector_store.id,
        file=file_handle,
    )


    
# Refund specialist: Handles refund requests
refund_agent = chat_client.as_agent(
    instructions="You process refund requests.",
    description="Agent that handles refund requests.",
    name="refund_agent",
    # In a real application, an agent can have multiple tools; here we keep it simple
    tools=[process_refund],
)
# Order/shipping specialist: Resolves delivery issues
order_agent = chat_client.as_agent(
    instructions="You handle order and shipping inquiries.",
    description="Agent that handles order tracking and shipping issues.",
    name="order_agent",
    # In a real application, an agent can have multiple tools; here we keep it simple
    tools=[FileSearchTool(vector_store_ids=[vector_store.id])],
)
# Return specialist: Handles return requests
return_agent = chat_client.as_agent(
    instructions="You manage product return requests.",
    description="Agent that handles return processing.",
    name="return_agent",
    tools=[FileSearchTool(vector_store_ids=[vector_store.id])],
    # In a real application, an agent can have multiple tools; here we keep it simple
    
)

    To Get format in certain structure, you can use structured output in orchestrating Agent or all other working Agents.

    Foundry IQ can be commonly used by multiple agents too. Please let us know if you had followed up queries on the same.

    Thank you.

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.