How to make the conversation finally back to the MAIN AGENT

[shinianzhihou]

Please read this first

• Have you read the docs?Agents SDK docs
• Have you searched for related issues? Others may have had similar requests

Yes

Question

Describe your question. Provide details if available.

My Agents is:

When i input a TASK to main_agent, it divide the main task into several tasks and deliever them to different agents. But the conversation(also the stream) always stop in some agent.

I want the other agents finally back to main_agents. Any functions or docs can help me solve it?

THANKS MY BEST MULTI-AGENTS-FRAMEWORK

[jackien1]

I had the same issue since I wanted to reduce costs by having different agents use cheaper or specialized models. Using agents as tools with a supervisor agent could work but I don't think the agent tools have access to message history. You could try using a while loop with a supervisor agent that hands off to other agents every turn and passes message history. To break the loop, use a LLM judge or some external code.

https://github.com/openai/openai-agents-python/blob/main/examples/agent_patterns/llm_as_a_judge.py

[rm-openai]

Can you share some sample code? In general there are a couple of ways to force this:

1. Prompting, telling the model to always pass back
2. Requiring a specific handoff tool to be used via tool_choice in ModelSettings

[shinianzhihou]

I had the same issue since I wanted to reduce costs by having different agents use cheaper or specialized models. Using agents as tools with a supervisor agent could work but I don't think the agent tools have access to message history. You could try using a while loop with a supervisor agent that hands off to other agents every turn and passes message history. To break the loop, use a LLM judge or some external code.

https://github.com/openai/openai-agents-python/blob/main/examples/agent_patterns/llm_as_a_judge.py

This helps me.
I use a judge to continue the loop conversation.

@dataclass
class EvaluationFeedback:
    feedback: str
    score: Literal["任务完成", "任务没有完成", "需要用户输入"]

[shinianzhihou]

Can you share some sample code? In general there are a couple of ways to force this:

1. Prompting, telling the model to always pass back
2. Requiring a specific handoff tool to be used via tool_choice in ModelSettings

I have tried the first way and it sometimes works.

The second way seems more elegant and could u plz provide some example codes?

I had the same issue since I wanted to reduce costs by having different agents use cheaper or specialized models. Using agents as tools with a supervisor agent could work but I don't think the agent tools have access to message history. You could try using a while loop with a supervisor agent that hands off to other agents every turn and passes message history. To break the loop, use a LLM judge or some external code.
https://github.com/openai/openai-agents-python/blob/main/examples/agent_patterns/llm_as_a_judge.py

This helps me. I use a judge to continue the loop conversation.

@DataClass
class EvaluationFeedback:
feedback: str
score: Literal["任务完成", "任务没有完成", "需要用户输入"]

use a judge also works but i cannot connect the agent with other agents.

[goooodle]

I had the same issue

[koakuma-chan]

@rm-openai Is there a better way to write this?

from agents import Agent

sponge_bob_agent = Agent(
    name="sponge_bob_agent",
    instructions="You are Sponge Bob.",
    handoffs=[patrick_agent], # Error: Agent Patrick is undefined
)

patrick_agent = Agent(
    name="patrick_agent",
    instructions="You are Patrick.",
    handoffs=[sponge_bob_agent],
)

from agents import Agent

sponge_bob_agent = Agent(
    name="sponge_bob_agent",
    instructions="You are Sponge Bob.",
    handoffs=[],
)

patrick_agent = Agent(
    name="patrick_agent",
    instructions="You are Patrick.",
    handoffs=[sponge_bob_agent],
)

sponge_bob_agent.handoffs.append(patrick_agent) # Awkward

[RoniFinTech]

Is there a good solution?

1. I tried with handoffs but I was facing an unreliable plan no matter how I changed the prompt of the supervisor agent instructions.
2. I tried the tools, but was facing hallucinations at different stages especially when I needed information that was generated in one agent to be used in the next agent.

any thoughts?