AzureAIClient with agent instructions

[cecheta]

Is it possible to create an agent from AzureAIClient and pass instructions to the agent?

When running the following code, I would expect an Azure AI agent to be created with instructions defined for the agent:

from agent_framework.azure import AzureAIClient
from azure.identity.aio import DefaultAzureCredential


async def main():
    agent = AzureAIClient(
        credential=DefaultAzureCredential(),
        agent_name="MyAgent",
    ).as_agent(instructions="You are a helpful assistant.")

    res = await agent.run("Hello.")
    print(res.text)


if __name__ == "__main__":
    import asyncio

    asyncio.run(main())

However, what actually happens is that the agent is not created with any instructions, and instead the "instructions" are passed as a system message when the agent is invoked. This has a downside that the same system message will be added on all subsequent invocations using the same thread.

It seems like it's possible to choose this if you use AzureAIProjectAgentProvider but not with AzureAIClient.

[cecheta]

To me, I would think that instructions added directly to the as_agent() call would be added to the agent's definition, while any instructions later passed to the agent while running would get converted to system messages.

    agent = AzureAIClient(
        credential=DefaultAzureCredential(),
        agent_name="MyAgent",
    ).as_agent(instructions="You are a helpful assistant.")  # <- Agent definition

    res = await agent.run("Hello.", options={"instructions": "Reply in uppercase."})  # <- System message

Currently, the two instructions get concatenated into a single system message, while the agent's instructions in its definition is empty.

[moonbox3]

Adding @dmytrostruk