Building AI Agents from Zero to Hero

Built AI agents with Strands SDK, Bedrock, and MCP from simple chatbots to an AWS Study Buddy with memory and tools.

Dineshraj Dhanapathy

Sharing, Learning, Growing Together

Published May 24, 2026

Last modified May 24, 2026

📖 Introduction

I participated in the AWSUG MDU April Skill Sprint, a 5-challenge journey to build AI agents from scratch using the Strands Agents SDK and Amazon Bedrock. Each challenge builds on the previous one, going from a simple chatbot to a full MCP-powered study assistant.

Here's my complete journey through all 5 challenges, including the code, architecture, errors I hit, and how I solved them.

🗺️ The Journey at a Glance

Challenge 1: My First Agent ⭐

Goal

Create a basic conversational agent using Strands SDK with a local LLM (Ollama).

Code

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
from strands import Agent

from strands.models.ollama import OllamaModel



model = OllamaModel(

    host="http://localhost:11434",

    model_id="llama3.2:3b"

)



agent = Agent(

    model=model,

    system_prompt="You are a helpful AI assistant.",

    tools=[]

)



response = agent("What is Python?")

print(f"🤖 Agent: {response}")

What I Learned

Strands SDK makes agent creation dead simple — 3 lines of config
OllamaModel connects to a local LLM (no cloud needed for testing)
The system_prompt shapes the agent's personality
tools=[] means no tools — just pure conversation

Output

1
2
🤖 Agent: Python is a high-level, interpreted programming language known for

its simplicity and readability...

Challenge 2: Adding Tools ⭐⭐

Goal

Give the agent superpowers with custom tools using Amazon Nova Pro via Bedrock.

Architecture

Key Code — The `@tool` Decorator

1
2
3
4
5
6
7
8
9
10
11
12
13
from strands.tools import tool



@tool

def weather(city: str) -\> str:

    """Get the current weather for a given city.

    

    Args:

        city: The name of the city to get weather for.

    """

    response = requests.get(f"https://wttr.in/{city}?format=j1")

    data = response.json()

    current = data["current_condition"][0]

    return f"Weather in {city}: {current['weatherDesc'][0]['value']}, {current['temp_C']}°C"

How Tool Selection Works

The agent reads the docstring and type annotations to understand:

When to use the tool (docstring describes purpose)
What to pass (type annotations define parameters)
What to expect back (return type)

You can see the agent's reasoning in \<thinking\> tags:

1
2
3
\<thinking\> I need to use the 'weather' tool to get the current weather

conditions in Chennai. \</thinking\>

Tool #1: weather

Output

1
2
3
4
5
🧮 Math: 42 * 17 = 714

🌤️ Weather: Chennai is thundery, 35°C, humidity 51%

🎂 Age: Someone born Jan 15, 2000 is 26 years old

📏 Conversion: 100 km = 62.14 miles

🌡️ Multi-tool: Delhi is 44°C (111.2°F) — used weather + unit_converter!

Key Insight

The agent chains multiple tools automatically. When asked "weather in Delhi in Fahrenheit", it calls weather first, then unit_converter without being told to.

Challenge 3: Persistent Memory ⭐⭐

Goal

Give the agent memory that survives program restarts using FAISS vector store.

Architecture

The Two Types of Memory

Type	Persists?	How
Conversation History	❌	Messages sent to model each turn
mem0_memory + FAISS	✅	Vectors stored on disk

Code

1
2
3
4
5
6
7
8
9
10
11
12
13
14
import os

os.environ["MEM0_LLM_MODEL"] = "amazon.nova-pro-v1:0"

os.environ["MEM0_EMBEDDER_MODEL"] = "amazon.titan-embed-text-v2:0"

os.environ["BYPASS_TOOL_CONSENT"] = "true"



from strands_tools import mem0_memory



agent = Agent(model=model, tools=[mem0_memory])



# Store

agent.tool.mem0_memory(action="store", content="My name is DD", user_id="dd_user")



# Retrieve

agent.tool.mem0_memory(action="retrieve", query="name", user_id="dd_user")

🔥 Errors I Hit (and Fixes)

Error	Root Cause	Fix
`No module named 'strands_tools'`	Wrong package name	`pip install strands-agents-tools`
`No module named 'mem0'`	Missing dependency	`pip install "strands-agents-tools[mem0-memory]"`
`No module named 'opensearchpy'`	Missing dependency	`pip install opensearch-py`
`Unsupported vector store provider: faiss`	mem0ai too new (1.x)	`pip install "mem0ai\>=0.1.99,\<1.0.0"`
`Unsupported embedding provider: aws_bedrock`	Same version issue	Same fix as above
Authorization error	mem0 defaults to Claude Haiku internally	Set `MEM0_LLM_MODEL` env var

output :

Critical Lesson

> mem0ai must be \<1.0.0. Version 1.x removed FAISS and aws_bedrock support. Always install with: pip install "strands-agents-tools[mem0-memory]"

Challenge 4: The Full Agent ⭐⭐⭐

Goal

Combine everything – tools, memory and streaming – in an interactive chat.

Architecture

Key Innovation: Simple File-Based Memory

After fighting mem0ai version issues, I replaced it with a dead-simple JSON approach:

1
2
3
4
5
6
7
8
9
10
11
12
13
@tool

def remember(key: str, value: str) -\> str:

    """Store info to persistent memory."""

    memory = json.load(open("memory.json"))

    memory[key] = value

    json.dump(memory, open("memory.json", "w"))

    return f"✅ Stored: {key} = {value}"



@tool

def recall(key: str) -\> str:

    """Retrieve from memory. Use key='all' for everything."""

    memory = json.load(open("memory.json"))

    return memory.get(key, "Not found")

Why this works better:

Zero dependencies
No version conflicts
No auth errors
Always works
Persists across restarts

Streaming Callback

1
2
3
4
5
def streaming_callback(**kwargs):

    if "data" in kwargs:

        print(kwargs["data"], end="", flush=True)

    if "current_tool_use" in kwargs and kwargs["current_tool_use"].get("name"):

        print(f"\n🔧 Using tool: {kwargs['current_tool_use']['name']}")

> Important: Use **kwargs, NOT positional arguments. The SDK passes everything as keyword args.

output

Demo Conversation

1
2
3
4
5
6
7
8
9
10
11
12
13
14
🧑 You: Remember my name is DD and I'm from Tanjore

🤖 Buddy: 🔧 Using tool: remember

🔧 Using tool: remember

✅ Stored! Your name is DD and you're from Tanjore! 🏛️



🧑 You: What's the weather in my city?

🤖 Buddy: 🔧 Using tool: recall

🔧 Using tool: weather

Weather in Tanjore: Sunny, 38°C ☀️



🧑 You: How old is someone born 2002-03-15? Also 365 * 24?

🤖 Buddy: 🔧 Using tool: age_calculator

🔧 Using tool: calculator

24 years old! And 365 × 24 = 8,760 hours in a year! 🧮

Challenge 5: MCP-Powered Innovation

Goal

Build a creative agent using MCP (Model Context Protocol) that solves a real problem.

My Project: AWS Study Buddy 📚

An interactive AWS tutor that:

📚 Searches real AWS documentation via MCP
🧠 Remembers what you've studied
📝 Quizzes you on learned topics
🧮 Calculates AWS costs
⚡ Streams responses in real time

more details about MCP project: Building an AWS Study Buddy with MCP + Strands Agents SDK | AWS Builder Center

Architecture

MCP Connection Code

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
from strands.tools.mcp import MCPClient

from mcp import StdioServerParameters, stdio_client



aws_docs_mcp = MCPClient(

    lambda: stdio_client(

        StdioServerParameters(

            command="uvx",

            args=["awslabs.aws-documentation-mcp-server@latest"],

            env={**os.environ, "FASTMCP_LOG_LEVEL": "ERROR"}

        )

    )

)



with aws_docs_mcp:

    mcp_tools = aws_docs_mcp.list_tools_sync()

    all_tools = mcp_tools + [save_note, get_notes, record_quiz_score, calculator]

    agent = Agent(model=model, tools=all_tools)

Learning Flow

output :

Demo

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
📚 AWS Study Buddy — Learn AWS with Official Docs + Memory



🧑 You: Explain Lambda in simple terms



🤖 Study Buddy:

🔧 Using tool: search_documentation

🔧 Using tool: read_documentation



AWS Lambda is serverless compute! ⚡

• Run code without managing servers

• Pay only for compute time used

• Triggers: API Gateway, S3, DynamoDB, etc.

• Max execution: 15 minutes

• Supports Python, Node.js, Java, Go, .NET



🔧 Using tool: save_note

📝 Note saved for 'Lambda'! You've now studied 1 topic.



🧑 You: How much for 1 million Lambda requests?



🤖 Study Buddy:

🔧 Using tool: calculator

Lambda pricing: $0.20 per 1M requests

1,000,000 × $0.0000002 = $0.20/month 💰

Plus compute time based on memory and duration.



🧑 You: Quiz me!



🤖 Study Buddy:

🔧 Using tool: get_notes



Q1: What's the max execution time for Lambda?

Q2: Name 3 Lambda triggers.

Q3: Do you manage servers with Lambda?

📊 Summary: Skills Progression

Challenge	New Skill	Model	Key Feature
1	Agent basics	Ollama (local)	System prompt, simple Q&A
2	`@tool` decorator	Nova Pro (Bedrock)	Weather, age calc, unit converter
3	Persistent memory	Nova Pro + Titan Embed	FAISS/JSON storage across sessions
4	Streaming + multi-tool	Nova Pro	Callback handler, interactive chat
5	MCP protocol	Nova Pro + MCP Server	AWS docs search, quiz system

🔥 Top Errors & Lessons

#	Error	Lesson
1	`pip install strands-tools` fails	Package is `strands-agents-tools`
2	`mem0ai` 1.x breaks everything	Pin to `\<1.0.0` or use `[mem0-memory]` extra
3	`streaming_callback(event)` fails	Use `**kwargs` not positional args
4	mem0 authorization errors	Set `MEM0_LLM_MODEL` env var or use simple JSON memory
5	MCP server not found	Install `uv` for `uvx` command

🛠️ Full Tech Stack

Component	Technology
Agent Framework	Strands Agents SDK
LLM	Amazon Nova Pro v1 (Bedrock)
Embeddings	Amazon Titan Embed Text V2
MCP Server	awslabs.aws-documentation-mcp-server
Memory	JSON file (simple) / FAISS (advanced)
Weather API	wttr.in
Language	Python 3.12
Region	us-east-1

🏗️ Project Structure

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
may-skill-sprint/

├── challenge-1-first-agent/

│   └── starter.py          # Basic agent with Ollama

├── challenge-2-tools/

│   └── starter.py          # Custom tools + Bedrock

├── challenge-3-memory/

│   ├── starter.py          # mem0_memory + FAISS

│   ├── verify_memory.py    # Persistence test

│   └── debug_mem0.py       # Debug script

├── challenge-4-full-agent/

│   ├── starter.py          # Full interactive agent

│   └── memory.json         # Persistent memory store

├── challenge-5-mcp-agent/

    ├── starter.py          # MCP-powered study buddy

    └── study_notes.json    # Study progress

🎓 Final Thoughts

Building AI agents is easier than you think but getting the dependencies right is harder than you'd expect. The Strands SDK itself is elegant and simple. The pain points are all in the ecosystem (version conflicts, auth issues, package naming).

My biggest takeaway: start simple, add complexity gradually. Challenge 1 → 5 is the perfect progression. And when something complex breaks (like mem0), don't be afraid to replace it with something simpler that just works.

The future of AI development is agents + tools + memory + MCP. This sprint gave me hands-on experience with all four.

GitHub link : https://github.com/dineshrajdhanapathyDD/Building-AI-Agents-from-Zero-to-Hero.git

Built with ❤️ by DD for the AWSUG MDU April Skill Sprint 2026

Thank you for taking the time to read my article! If you found it helpful, feel free to like, share, and drop your thoughts in the comments; I’d love to hear from you.

# amazon-bedrock # strands-agents # python # amazon-nova # mcp

Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.

Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article

Builder Center

Building AI Agents from Zero to Hero

Dineshraj Dhanapathy

📖 Introduction

🗺️ The Journey at a Glance

Challenge 1: My First Agent ⭐

Goal

Code

What I Learned

Output

Challenge 2: Adding Tools ⭐⭐

Goal

Architecture

Key Code — The @tool Decorator

How Tool Selection Works

Output

Key Insight

Challenge 3: Persistent Memory ⭐⭐

Goal

Architecture

The Two Types of Memory

Code

🔥 Errors I Hit (and Fixes)

output :

Critical Lesson

Challenge 4: The Full Agent ⭐⭐⭐

Goal

Architecture

Key Innovation: Simple File-Based Memory

Streaming Callback

output

Demo Conversation

Challenge 5: MCP-Powered Innovation

Goal

My Project: AWS Study Buddy 📚

Architecture

MCP Connection Code

Learning Flow

output :

Demo

📊 Summary: Skills Progression

🔥 Top Errors & Lessons

🛠️ Full Tech Stack

🏗️ Project Structure

🎓 Final Thoughts

Key Code — The `@tool` Decorator