__Agents

Agent

• An Agent is a system that leverages AI model to interact with its environment in order to achieve a user-defined objective
• It combines reasoning, planning, and the execution of actions (often via external tools) to fulfill tasks
• Examples
  • Virtual Assistants like Siri, Alexa, Google Assistant
    • analyze context, initiate actions like setting reminders
  • Customer Service chatbots
    • answer questions, guide and troubleshoot, complete transactions
  • AI NPC in video game
    • respond contextually, adapt to player interactions
• Ref:
  • https://huggingface.co/learn/agents-course/unit1/what-are-agents
  • https://codingdidi.substack.com/p/agentic-ai-engineering-mastery-with

Tools

• An Action taken by Agent can involve use of multiple Tools to complete
• Examples of Tools
  • Web Search
  • Image Generation
  • Retrieval
  • API interface
• A Tool should contain
  • Textual description of function
  • Callable
  • Arguments with typings
  • Outputs with typings
• @tool decorator automatically provides a to_string() method which provides all this info

@tool
def calculator(a: int, b: int) -> int:
    """Multiply two integers."""
    return a * b
 
print(calculator.to_string())
# Tool Name: calculator, Description: Multiply two integers., Arguments: a: int, b: int, Outputs: int

• This description is injected in the system prompt

system_message="""You are an AI assistant designed to help users efficiently and accurately. Your primary goal is to provide helpful, precise, and clear responses.
 
You have access to the following tools:
Tool Name: calculator, Description: Multiply two integers., Arguments: a: int, b: int, Outputs: int
"""

Thought-Action-Observation cycle

• Agent follows continuous cycle of:
  • Thought: The LLM part of the Agent decides what the next step should be.
  • Action: The agent takes an action by calling the tools with the associated arguments.
  • Observation: The model reflects on the response from the tool.
• Agent frameworks inject this in system prompt

system_message="""You are an AI assistant designed to help users efficiently and accurately. Your primary goal is to provide helpful, precise, and clear responses.
 
You have access to the following tools:
Tool Name: calculator, Description: Multiply two integers., Arguments: a: int, b: int, Outputs: int
 
You should think step by step in order to fulfill the objective with a reasoning divided into Thought/Action/Observation steps that can be repeated multiple times if needed.
 
You should first reflect on the current situation using 'Thought: {your_thoughts}', then (if necessary), call a tool with the proper JSON formatting 'Action: {JSON_BLOB}', or print your final answer starting with the prefix 'Final Answer:'
"""

Memory Architectures

• https://huggingface.co/learn/context-course/unit6/agent-loop
• Production Agents uses various memory architectures
  • Short term scratch pad: Save working state
  • Episodic Memory: Recalling past sessions
  • Semantic Memory: For persistent knowledge
  • Retrieval Augmented approaches: Fetch relevant memories on demand
  • Compaction Strategies: Summarize Older Context

Context Management

• Real Code Agents implement
  • Compaction
  • Structured Note Taking
  • File-System Mediated Context
  • Intelligent Tool Selection

Sandboxing

• Agent has internal tools available to it
• These tools are sandboxed to prevent running arbitrary dangerous code
• Following are techniques
  • Path confinement: Verify the path is in workspace
  • Command Allowlist: Only allowed commands are run, Anything which mutates state (rm, mv) or reaches network (curl, wget) is rejected
  • Output Limits: Limit output from say DB to avoid reading too much data
  • Write-Off-by-default: Turn off writes by default
  • Exception Handling: Graceful handling of exceptions
  • Timeout: To avoid hanging network call, or long running commands
• Avoid global imports when running tools via Agent loop

Nano-Harness

• https://huggingface.co/learn/context-course/unit6/introduction
• Agentic Loop
  • Call LLM with task + message history
  • Parse model output as Python code
  • Execute Python (with tools available)
  • Collect stdout, stderr, exceptions
  • Append observation to message history
  • Repeat (max 50 steps)
  • Done when: final_answer() called or max steps reached

Prompting Techniques

• CoT (Chain-of-Thought)
• ReAct (Reasoning + Acting)

Feature	CoT	ReAct
Step-by-step logic	✅ Yes	✅ Yes
External tools	❌ No	✅ Yes (Actions + Observations)
Best suited for	Logic, math, internal tasks	Info-seeking, dynamic multi-step tasks