How do I control the tool-calling in Autogen?

[SrijitaRoy913]

I am working on a multi-agent framework, and one of the agents calls a tool to fetch the schema of tables. The agent ends up suggesting tool calls multiple times simultaneously instead of handling it in one parameter.

This is a sample of the output:

first_assistant (to User):

***** Suggested tool call (call_qQoQXF1VY1MAWzfTTvSx1CKs): fetch_relevant_ddl *****
Arguments:
{"table_list": ["products"]}

---

***** Suggested tool call (call_DDDrgYA5DwUya0IDBWbcKZ5w): fetch_relevant_ddl *****
Arguments:
{"table_list": ["orders"]}

---

---

However, it will be better if the tool-calling were to happen just once, like this:

first_assistant (to User):

***** Suggested tool call (call_qQoQXF1VY1MAWzfTTvSx1CKs): fetch_relevant_ddl *****
Arguments:
{"table_list": ["products", "orders"]}

---

---

Is there any way to make the calling deterministic other than telling it through the prompt? Are there any parameters on configuration that needs to be enabled while defining the tool?

[ekzhu]

There is currently no easy way to achieve this in v0.2. For deterministic tool calls, what's the scenario? Do you require another inference on the tool call result or just return the result of the tool call directly.

[SrijitaRoy913]

I'm currently using these libraries:
autogen-agentchat==0.2.38.
autogen==0.4.1
pyautogen==0.3.1

I'm using GPT 4o-mini model.

The agents are in a Sequential Workflow and are configured in this way:

1. Agent 1: Understands the business use-case and outputs a task plan.
2. Retrieves schema. Has a tool to retrieve.
3. Generates SQL
4. Executes SQL and retries if errors. Has a tool to execute.

Can this flow be optimised?

[ekzhu]

autogen-agentchat==0.2.38.
autogen==0.4.1
pyautogen==0.3.1

All these libraries are going to cause conflict, just keep autogen-agentchat

Agent 1: Understands the business use-case and outputs a task plan.
Retrieves schema. Has a tool to retrieve.
Generates SQL
Executes SQL and retries if errors. Has a tool to execute.

1. Can you perform step 2 ahead of time? E.g., put all relevant schema in the context/input?
2. Can you combine step 1, 3 and 4 as a single agent with iterative tool call?

[SrijitaRoy913]

Got it, I'll remove the other libraries...

1. That might not be possible because relevant schema is retrieved based on the user's question.
2. I can explore this, how to do iterative tool call?

Also, is there any way I can manage cache to speed up?

[ekzhu]

For 2. if you have a user proxy initiate conversation with the agent, the two-agent chat is iterative tool call.

Also, is there any way I can manage cache to speed up?

What do you mean? The cache is turned on by default.

[SuMiaoALi]

I had met this problem. I solved it through provide better prompts to tool_caller agent.

[ranga-tec]

Try this aproach https://github.com/[ranga-tec/Autogen-dynamic-functions](https://github.com/ranga-tec/Autogen-dynamic-functions)