Lesson 9: The Anatomy of an Agent

Topics Covered

• The Core Loop: The circular architecture of autonomous systems.
• Sense: Normalizing the messy world into tokens.
• Think: The cognitive engine (Planning and Routing).
• Guardrails: The critical safety layer (Input/Output checks).
• Act: The difference between "generation" and "side effects."

An AI Agent isn't just "smarter" software; it is a system designed to mirror biological cognition. To build one, we must replicate the fundamental phases of intelligent life: Sensing, Thinking, and Acting, with a critical layer of Safety.

1. Sense (Perception & Normalization)

Before an agent can act, it must understand the state of the world. Since LLMs only understand tokens (text/code), the "Sense" phase is entirely about converting raw signals into text.

• User Input: The prompt (e.g., "Refund this order").
• Digital State: Reading database rows, checking API status codes, or parsing email headers.
• Physical Sensors: Converting camera pixels or microphone audio into text descriptions.

The Engineering Goal: Create a unified "Context Object" that represents the current reality.

2. Think (The Cognitive Engine)

This is the Agent's "Brain." It takes the Context Object and decides what to do next using the ReACT pattern.

• Plan: Break the user's goal into steps. ("First verify the order, then check the policy, then process the refund.")
• Route: Decide which specialized tool is needed. ("This is a billing issue, not technical support.")
• Memory Access: Retrieve user preferences from the Vector Database.

3. Guardrails (The Safety Layer)

An autonomous agent needs a conscience. Guardrails are programmable rules that sit between the Brain and the World.

• Input Rails (Before the LLM):
  • PII Scrubbing: Automatically redaction of credit card numbers or SSNs before they reach the model.
  • Jailbreak Detection: Blocking prompts that try to trick the agent (e.g., "Ignore all previous instructions").
• Output Rails (Before the Action):
  • Hallucination Check: Does the answer cite a source that actually exists?
  • Forbidden Topics: Ensuring the "Travel Agent" doesn't give medical advice.
  • Syntax Validation: Ensuring the SQL query generated by the model doesn't contain DROP TABLE commands.

4. Act (The Hands)

This is where the agent leaves the safety of the chat window and touches the real world.

• Digital Actions: Executing a SQL query, sending a POST request to Stripe, or posting to Slack.
• Physical Actions: Sending a signal to a robotic arm.

Critical Concept: Side Effects In Generative AI, the output is just pixels on a screen. In Agentic AI, the output is a Side Effect (money moved, data deleted). This requires strict permission layers—agents should often have "Read" access by default but require human confirmation for "Write" actions.

5. Learn (The Feedback Loop)

How does an agent get better?

1. Runtime Observation: The agent acts -> The tool returns an error -> The agent reads the error -> The agent tries a different parameter (Self-Correction).
2. Offline Fine-Tuning (RLHF): Engineers review logs. Successful paths are reinforced; failed paths are penalized.

6. Applied Anatomy: The "Refund Bot"

Phase	Action	System Component
Sense	Ingest email: "Where is my refund?"	NLP Parser
Think	Lookup User ID. Retrieve Order #123.	RAG + LLM Logic
Guardrail	Check: Is the refund amount < $50 (Auto-approve) or > $50 (Manager Approval)?	Python Policy Logic
Act	Call StripeAPI.issueRefund(123).	Tool / API Client
Learn	API returns "Success." Close ticket.	Feedback Loop