In an AI Agent flow, in order to complete a task that requires executing 2 times the same action (GetClinic) and one time another action ( GetClinicByInsurance) , how many LLM calls are required, at minimum, to process the intentions and generate the final conclusion?
A. 1 B. 2 C. 3 D. 4
The correct answer is A. 1.
In an AI Agent flow in OutSystems Developer Cloud (ODC), the number of LLM (Large Language Model) calls required to process the intentions and generate the final conclusion is one. This is because the AI Agent does not call the LLM for each individual server action execution. Instead, the LLM processes the user's intent, determines the sequence of actions to perform (in this case, GetClinic twice and GetClinicByInsurance once), and generates a single response based on the specified task or system prompt. The actual number of server actions executed is decoupled from the number of LLM calls.
GetClinic
GetClinicByInsurance
This behavior is consistent with the information provided in the OutSystems documentation, where an Agent Execution is considered as a single CallAgent action execution, even if the Agent invokes multiple server actions or repeats the same action during its reasoning process AI agent actions.
CallAgent
Hi André,
Please don't spam the forum like that.
Hello,
I think, given the question's limited context, the answer should not be 1 and our lovely OutSystems Neo has provided a misleading answer. It may be a problem with the way the question was phrased here.
First, there are two concepts in the answer: the Agent and the Model (LLM).And according to the documentation (cited by Neo), this is the agent's flow when there are actions available:As you can see by the flow, whenever the agent has actions configured, after the initial call to the Model, it decides what actions are required to be executed, and after executing them, there is another call to the model so it can decide if a new turn is required or if it has the final answer.The documentation also states that the Model can decide if it is required to call a single action, a sequence of actions or even call the same action multiple times.The question as it is phrased does not have enough information for us to know what will be the model decision. It can call one, than decide to call another one, than decide to call the final one. Or it can decide to call all of them in sequence. Or other combinations.In this scenario where there will be 3 actions being called, the minimum number of calls to the LLM would be two (2), not one (1). We call the Model once at start, it can answer saying it is necessary to call the 3 actions in sequence, we call the Model to give the actions response back to it and it can decide there is no more need to reason and it has the final answer.It’s important to notice that at minimum, whenever there is a call to an action, sequence of actions or the same action multiple times as a response to the Model answer, an extra call to the Model will be made after the actions execute, and the model may decide that it does not have yet the final answer, and need to be called again, or may decided it has and return the final answer.That’s one of the reasons why we have the call condition in the agent, so that we can define a condition to stop the loop in case the agent gets stuck.I would love to read your thoughts on this.
Ok, it seems I made a tiiiiinytinylittle mistake in my answer. It seems the "Send Action Result back to the Model" does not count as a call to the model.So... Sorry Neo, you were right. :(The Minimum number of calls, without knowing if the actions are being called all at once or one every reason-act loop is indeed 1, not 2.
This is a conceptual question about AI Agent flows in OutSystems ODC.
In an AI Agent flow, each tool call (action execution) requires the LLM to:
The agent loop works as: LLM decides → tool executes → LLM evaluates result → decides next step → ... → final answer. Each iteration of this loop requires one LLM call, and the final synthesis also requires one.
The answer is D. 4