Enhancing AI Agents: Adding Instructions, Tasks and Memory
– Overview: AI Agents: A Comprehensive Introduction for Developers
– Step 1: How To Define an AI Agent Persona by Tweaking LLM Prompts
– Step 2: Enhancing AI Agents: Adding Instructions, Tasks and Memory
– Step 3: Enhancing AI Agents: Implementing Reasoning Through Prompt Engineering
– Step 4: How To Add Persistence and Long-Term Memory to AI Agents
– Step 5: How To Add RAG to AI Agents for Contextual Understanding
– Step 6: How To Add Tool Support to AI Agents for Performing Actions
In our previous article on AI agents, we explored how to create a basic AI agent with a persona using system prompts. Now, we’ll delve into three crucial enhancements that make our agents more sophisticated and effective: instructions, tasks, and conversation memory.
Understanding Instructions vs. Persona
While a persona defines an agent’s core traits and role (like being a financial advisor), instructions provide specific guidelines for how the agent should operate. Think of the persona as like an employee’s job title and basic qualifications, while instructions are like the detailed operational manual they follow.
Let’s look at how instructions are implemented in our enhanced agent class:
|
1 2 3 4 5 6 7 8 9 |
@property def instruction(self) -> str: return self._instruction @instruction.setter def instruction(self, value: str): self._instruction = value # Reset conversation history when instruction changes self._conversation_history = [] |
The separation between persona and instruction allows for more flexible and precise control over the agent’s behavior. For example, a financial advisor agent might have:
- Persona: Core identity as a financial expert.
- Instructions: Specific guidelines about never recommending individual stocks, always including disclaimers, etc.
Understanding Tasks
While personas define the role and instructions provide operational guidelines, tasks represent the specific queries or actions requested of the agent.
|
1 2 3 4 5 6 7 |
@property def task(self) -> str: return self._task @task.setter def task(self, value: str): self._task = value |
In our implementation, tasks are handled distinctly from instructions, allowing for clearer separation of concerns:
- Instructions: Long-term guidelines that persist across conversations.
- Tasks: Immediate requests that need to be addressed.
Tasks work together with instructions to form the complete user-side context in the prompt structure. Here’s how they combine in the message chain:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 |
def _build_messages(self) -> List[Dict[str, str]]: """Build the messages list for the API call including system, instruction and history.""" messages = [{"role": "system", "content": self.persona}] if self.instruction: messages.append({"role": "user", "content": f"Global Instruction: {self.instruction}"}) # Add conversation history messages.extend(self.history) # Add current task if exists if self.task: messages.append({"role": "user", "content": f"Current Task: {self.task}"}) return messages |
This structure ensures that:
- Instructions provide the operational framework.
- Tasks represent the immediate objective.
- The agent maintains awareness of both the general guidelines and specific requests.
Task Processing Flow
When a task is executed, the following process takes place:
- The agent first considers its persona (system role).
- The agent then applies any global instructions.
- Finally, it processes the specific task.
- The response is stored in the conversation history.
This layered approach allows the agent to:
- Maintain its professional identity (persona).
- Follow consistent guidelines (instructions).
- Address specific needs (tasks).
- Build context over time (memory).
When to Use Instructions vs. Tasks
Understanding the distinction between instructions and tasks is fundamental to designing effective AI agents. Instructions serve as the foundational framework that guides an agent’s behavior across all interactions. Think of instructions as the operating manual for the agent — they define broad behavioral guidelines, establish persistent rules and boundaries, and set methodologies for handling various situations. Like a company’s standard operating procedures, instructions remain consistent across multiple interactions, ensuring the agent maintains reliable and has predictable behavior patterns.
While instructions provide the overarching framework, tasks are the practical implementation of that framework in response to specific user requests.
Tasks, on the other hand, represent the specific actions an agent needs to perform at any given moment. They are the concrete, immediate objectives that change based on user needs and requirements. While instructions provide the overarching framework, tasks are the practical implementation of that framework in response to specific user requests. Tasks must always operate within the boundaries set by the instructions, much like how an employee performs specific assignments while adhering to company policies.
Building Conversation Memory
One of the most significant enhancements is the addition of conversation memory. This allows our agent to maintain context throughout an interaction, much like how a human professional would remember previous parts of a conversation with a client.
The implementation uses a history property to track the conversation history:
|
1 2 3 |
@property def history(self) -> List[Dict[str, str]]: return self._history |
The history, which acts as a short-term memory system, serves several purposes:
- Contextual Understanding: The agent can reference previous interactions.
- Consistency: Responses remain coherent across multiple exchanges.
- Learning: The agent can build upon previous discussions.
Constructing Message Chains
The enhanced agent uses a sophisticated message-building system that combines persona, instructions, and conversation history.
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 |
def execute(self, task: str = None) -> str: """ Execute the agent with the given task or use existing task if none provided. Updates conversation history with the interaction. """ if task: self.task = task if not self._api_key: return "API key not found. Please set the OPENAI_API_KEY environment variable." client = OpenAI(api_key=self._api_key) messages = self._build_messages() try: response = client.chat.completions.create( model=self._model, messages=messages ) # Extract response content response_content = response.choices[0].message.content # Update history with the interaction if self.task: self._history.append({"role": "user", "content": self.task}) self._history.append({"role": "assistant", "content": response_content}) # Clear current task after execution self._task = "" return response_content except Exception as e: return f"An error occurred: {str(e)}" |
This structured approach ensures that:
- The persona remains the foundational context.
- Instructions are clearly acknowledged.
- Previous conversations provide additional context.
- The current task is properly framed within the larger context.
Practical Implementation
Let’s look at a practical example using a financial advisor agent:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 |
# Create a financial advisor agent financial_advisor = Agent("FinancialAdvisorBot") # Set the persona (core traits and role) financial_advisor.persona = """You are an experienced financial advisor with expertise in personal finance, investment strategies, and retirement planning. Always maintain a professional and ethical approach to financial advice.""" # Set the global instruction (broader guidelines) financial_advisor.instruction = """When providing financial advice: 1. Always emphasize the importance of individual circumstances and risk tolerance 2. Never recommend specific stocks or make promises about returns 3. Include disclaimers about consulting licensed professionals when appropriate 4. Break down complex concepts into simple, actionable steps 5. Consider both short-term and long-term implications in your advice""" |
The agent now maintains context across multiple interactions while following specific guidelines:
|
1 2 3 4 5 |
task1 = "What are some key considerations for planning retirement in your 30s?" response1 = financial_advisor.execute(task1) task2 = "Based on those retirement considerations, what should my monthly savings target be?" response2 = financial_advisor.execute(task2) |
In this example, the second response would take into account the context from the first interaction, providing more relevant and connected advice.
The entire source code for the framework is available in a GitHub repository.
Benefits of Enhanced Agents
These enhancements bring several key advantages:
- Improved Consistency: The combination of persona and instructions ensures that responses are consistently aligned with the agent’s role and guidelines.
- Better Context Management: Conversation memory allows for more natural and coherent interactions across multiple exchanges.
- Flexible Behavior Control: Separating a persona from instructions makes it easier to modify how the agent operates without changing its core identity.
- Enhanced User Experience: Users receive more contextually appropriate responses that build upon previous interactions.
Limitations and Next Steps
These enhancements represent a significant step forward in creating more sophisticated AI agents. However, there’s still room for further improvements, such as:
- Long-term memory storage and retrieval;
- Task-specific instruction sets;
- Dynamic persona adaptation based on user needs and
- Enhanced error handling and edge case management.
By continuing to refine and enhance these components, we can create AI agents that are increasingly capable of handling complex, multi-step interactions while maintaining consistency and contextual awareness.
The combination of persona, instructions, and memory brings us closer to creating AI agents that can truly serve as effective digital assistants across a wide range of professional domains.
In the next part of this series, we will add reasoning to agents based on proven prompt engineering techniques. Stay tuned!
