Agents not cleaned up Ai foundry – Duplicate agent instances created on each run in Multi-Agent Accelerator

[UtkarshMishra-Microsoft]

Technical Feedback
We are encountering a persistent issue in the Multi-Agent Custom Automation Engine Solution Accelerator where multiple agent instances are being created and stored under the "Agents" tab in the Azure AI Foundry portal. These agent instances are not being deleted automatically after each application run or upon shutdown, resulting in agent duplication across sessions. Over time, this leads to an unnecessarily large number of agents, consuming resources and cluttering the Foundry interface.

There appears to be no built-in mechanism to clean up agents after they are no longer in use. Even though the code uses an in-memory cache during execution, it does not prevent new agents from being created on every run, and it does not remove the old ones from the Foundry portal.

Desired Outcome
The expected behavior is that the Foundry project should not retain redundant agent instances. Ideally, the system should implement one of the following:

Automatically delete all agents when the application shuts down or finishes its run,
OR

Preserve only one instance of each agent type per user (reusing them across sessions if applicable).

This would ensure that the Agents tab (accessible from the Foundry project > Foundry Portal > Agents in the left-hand sidebar) remains clean and manageable. The critical goal is to prevent agents from being recreated repeatedly without cleanup, which currently leads to potential quota issues and operational confusion.

Current Workaround

Describe the current workaround.
Looking for some workaround so agents gets cleanup after using app or shutting down the app .

Please help with this issue!!!

[guygregory]

Hi @UtkarshMishra-Microsoft, this behavior is by design, as the agents are designed to be persistent.

If you want to clean-up after the agent is used, then you can simply use the delete_agent() method, as described in the documentation

# Delete the agent when done
project_client.agents.delete_agent(agent.id)
print("Deleted agent")

An alternative approach is to check for the existence of the agent each time, and only create a new one if it doesn't already exist. There's a good example of how to do this in @nicholasdbrady's demo here. I've found this method works well for interactive chat apps, where the user may terminate the app unexpectedly.

AGENT_NAME = "my-enterprise-agent"
found_agent = None
all_agents_list = project_client.agents.list_agents().data
for a in all_agents_list:
    if a.name == AGENT_NAME:
        found_agent = a
        break

model_name = os.environ.get("MODEL_DEPLOYMENT_NAME", "gpt-4o")
instructions = (
    "You are a helpful enterprise assistant at Contoso. "
    f"Today's date is {pydatetime.now().strftime('%A, %b %d, %Y, %I:%M %p')}. "
    "You have access to hr documents in file_search, the grounding engine from bing and custom python functions like fetch_weather, "
    "fetch_stock_price, send_email, etc. Provide well-structured, concise, and professional answers."
)

if found_agent:
    # Update the existing agent to use new tools
    agent = project_client.agents.update_agent(
        assistant_id=found_agent.id,
        model=found_agent.model,
        instructions=found_agent.instructions,
        toolset=toolset,
    )
    print(f"reusing agent > {agent.name} (id: {agent.id})")
else:
    agent = project_client.agents.create_agent(
        model=model_name,
        name=AGENT_NAME,
        instructions=instructions,
        toolset=toolset
    )
    print(f"creating agent > {agent.name} (id: {agent.id})")

[UtkarshMishra-Microsoft]

Thanks @guygregory for the insights. I’d like to confirm how thread handling is implemented, particularly in scenarios where threads are not required—as we are deleting the agents from foundry.

[UtkarshMishra-Microsoft]

I'd like to confirm how thread handling is implemented, especially in cases where threads are no longer needed—such as when we're deleting agents from Foundry. It seems that thread creation is an automatic feature of Semantic Kernel when an agent is created. However, we're unsure how to delete the associated threads, and we haven't been able to figure it out so far. Do you have any insights on this?