AgentLogic

AgentLogic module is responsible for handling a single iteration of the logic part of the agent, i.e. without executing any external tool. In this role AgentLogic is the sole place of full agent where we interact with the language model.

AgentLogic object is initialized as:

def __init__(
        self,
        context: AgentContext,
        model: LLM,
        prompt_strategy: PromptStrategy,
        stop_word: Optional[str] = None,
    ) -> None

A generic AgentLogic class brings the run method that represents this single iteration. This function is:

def run(
        self, 
        instruction: str, 
        short_memory: ShortMemory
    ) -> str:

    # History
    history = short_memory.get_all()

    # Assemble the messages
    msgs = self._prompt_strategy.render(
        instruction, 
        history
    )

    # Run the LLM
    output = self._llm.run(
        msgs, 
        self._stop_word
    )

    # Process the response
    response = self.process_response(output)

    return response

AgentLogic uses language model object defined in the llm module. This module is discussed in details in the Language and Embedding Models Abstraction section.

Example Agent Logic Instantiation

agent_context = AgentContext(
        agent_name="TestAgent",
        agent_persona="Researcher always following scientific method",
    )

llm_factory = LLMFactory()
llm_model = llm_factory.create("ollama|gemma3")

template = FileTemplate(
        agent_context.model_dump(), 
        "ReAct.template"
    )

strategy = ReactPrompt(
        system_template=template,
        tool_dictionary={},
    )

sm = ShortMemory(capacity=5)

agent_logic = GenericAgentLogic(
        context=agent_context,
        model=llm_model,
        prompt_strategy=strategy,
        stop_word="<PAUSE>",
    )

INSTRUCTION = "Why is the sky blue?"
result = agent_logic.run(
        instruction=INSTRUCTION, 
        short_memory=sm
    )

In this example we create a simple agent with the ReAct prompt strategy. The agent is initialized with the AgentContext, LLM, PromptStrategy and the stop word.

The agent's logic is then run with the instruction and the short memory. The result of the execution is returned as a string.

The setup of the agent logic may seem cumbersome as it involves number of objects and steps. As such, in Elemental it is included in the full agent initialization. However, all components are exposed for user to build more custom solution.

Core Components

Memory