Why Agents Matter
Think about how you solve a complex problem. You do not just think about it, you:
- →Reason about what to do next
- →Act by looking something up, running a calculation, or asking someone
- →Observe the result
- →Repeat until the problem is solved
This is exactly what ReAct agents do. They are AI systems that can plan, use tools, learn from feedback, and iterate toward a solution.
The ReAct Loop
ReAct: Reason, Act, Observe
The core loop of every AI agent
Thought
Thought: The user wants to know the current weather in Paris and whether they need an umbrella. I need to check the weather API first.💡 The thought step makes the agent's reasoning explicit and debuggable.
Agents vs Chains: When to Use What
Chains vs Agents
← Scroll horizontally →
What makes ReAct agents more capable than pure CoT reasoning?
Tools: The Agent's Hands
Safety and Guardrails
- →Tool whitelisting, Only give agents access to tools they need. A customer support agent does not need file system access.
- →Confirmation for destructive actions, Require human approval before sending emails, making purchases, or deleting data.
- →Iteration limits, Set a maximum number of ReAct loops (typically 5-10) to prevent infinite loops.
- →Output validation, Check the agent's final answer against the observations to detect hallucination.
- →Cost controls, Each loop is an API call. Set budget limits to prevent runaway costs.
Test Your Understanding
ReAct Agents Knowledge Check
0 / 5 answered
Q1.What does ReAct stand for?
5 questions remaining
What's Next
You now understand the ReAct framework. In the next workshop, you will build your own AI agent with tool use and safety guardrails.
- →AI Agent Architecture Patterns, The 5 workflow patterns for building effective agents
Continue to the workshop: Build an AI Agent with Safety to create an autonomous agent.
Workshop Goal
By the end, you will have built an agent that:
- →Takes a user question and plans a series of actions
- →Uses 3 tools: web search, calculator, and document lookup
- →Requires human confirmation for any action that modifies data
- →Stops after 7 iterations or $0.50 in API costs
- →Validates its final answer against observed evidence
Step 1: Define the Agent Architecture
Building the Agent Step by Step
From system prompt to safety layer
The System Prompt
You are a research assistant. You help users find accurate information.
Available tools:
1. search_web(query) - Search the internet. Use for current events or facts.
2. calculator(expression) - Evaluate math expressions. Use for any calculation.
3. lookup_docs(query) - Search internal documents. Use for company-specific info.
Rules:
- Always explain your reasoning before acting
- Use the minimum number of steps
- Never guess when you can look up
- If uncertain after 5 steps, say so honestly💡 Rules in the system prompt act as behavioral guardrails.
Step 2: Implement Safety Layers
Safety Layer Implementation
← Scroll horizontally →
Why is human-in-the-loop important for AI agents?
Step 3: Testing Your Agent
Agent Testing Strategy
Verify behavior before deployment
Happy Path Testing
Test: 'What is the current population of Tokyo?'
Expected: search_web -> extract number -> answer
Actual: Verify tool selection and answer accuracy💡 Start with simple cases to verify the basic loop works.
Common Agent Anti-Patterns
Test Your Understanding
AI Agent Workshop Check
0 / 5 answered
Q1.What is the most common source of agent failures in production?
5 questions remaining
What's Next
You have built an AI agent with proper safety guardrails. In the next module, you will explore AI image generation, understanding diffusion models and learning to prompt for visual content.
Continue to AI Image Generation and Diffusion Models to enter the world of visual AI.