Strategy Is Actually Easy (If You Have AI Agents)

Here is the 60-second demo video of AI STRATEGIC PLANNER

This is a submission for the Google AI Agents Writing Challenge: Learning Reflections

---

Your Story + URL + AI Agents = Your Own BMC

In this post, I’ll share how the 5-Day AI Agents Intensive with Google and Kaggle helped me design a multi-agent system that turns a founder’s story and website into a personalized Business Model Canvas (BMC).

---

0. Who I Am and Why This Course Mattered

I’m not a full-time software engineer.

I work at a regional financial institution in Japan, supporting small and medium-sized businesses (SMEs) with:

• Business planning
• Subsidy and grant applications
• Long-term strategy discussions

In other words, I live in a world full of business plans and frameworks that many founders quietly dislike:

• “Please fill out this 3-year revenue plan.”
• “Please describe your competitive advantage in this format.”
• “Please upload a full business plan using this template.”

Strategic planning often feels like a high-friction, jargon-heavy domain, even for smart, experienced founders.

Before the 5-Day AI Agents Intensive, my mental model of AI was simple:

“LLMs are smart assistants that help me write, summarize, and answer questions.”

I had used generative AI for writing and analysis.

But “agentic systems” still felt abstract:

• How do you go from a single LLM call…
• …to a system that plans, uses tools, remembers, and hands work off between components?

This course gave me:

• A vocabulary for agents
• Architectural patterns I can reason about
• And most importantly, the confidence to design a multi-agent system for something I care deeply about:

Helping SME owners turn messy ideas into structured, actionable business plans.

---

1. What “Agents” Meant to Me Before the Course

Before Day 1, “AI agent” in my head roughly meant:

“Some wrapper around an LLM that takes a goal and then magically does stuff.”

I had seen frameworks and buzzwords:

• “Agents that use tools”
• “Autonomous AI workflows”
• “Multi-agent orchestration”

…but without a solid mental model, it was hard to go beyond copy–paste tutorials.

My default pattern was:

• One LLM
• One big prompt
• A few follow-up messages
• If something broke: tweak the prompt, try again, and hope

The 5-Day AI Agents Intensive basically said:

“Stop treating agents as bigger prompts.

Start treating them as systems.”

That mental shift is the single most important thing I’m taking away.

---

2. Concepts That Changed How I Think

Across the five days, a few ideas reshaped my understanding of agents.

2.1 Agents ≠ Chatbots

Day 1 clarified something I thought I knew:

A chatbot replies.

An agent does things.

It sounds obvious, but the course grounded it in architecture:

• Agents receive goals, not just messages
• They can select tools, call APIs, and update state
• They are designed to operate over time, not just in a single request–response

Once I saw agents as goal-driven systems, my expectations changed:

• I stopped expecting perfect one-shot answers
• I started thinking in terms of robust workflows

2.2 Tools and MCP: Interoperability by Design

Day 2 focused on tools and the Model Context Protocol (MCP).

Before that, my approach was:

“Glue services together myself and feed the results into the LLM.”

MCP offered a more principled way to:

• Declare tools in a structured format
• Expose external systems (files, APIs, note stores)
• Let agents discover and call these tools in a consistent way

For my use case, this meant I could:

• Treat something like Obsidian (for notes & logs) as a first-class tool
• Let agents read/write context instead of hard-coding everything into prompts

This is the moment when “multi-agent SME advisor” stopped being a slide idea and became implementable architecture.

2.3 Context Engineering: Memory ≠ Full History

Day 3’s “context engineering” hit me hard.

My naive approach was:

“Throw the entire conversation history into the prompt and call it memory.”

The course suggested a more disciplined view:

• Short-term vs long-term context
• What to summarize, and how
• What to store structurally
• What belongs in external tools vs in the prompt

For SME business planning, this is crucial:

• The founder’s story
• Key assumptions and constraints (cash, team, time)
• External data (market info, benchmarks)

You don’t want to shove all of that into a single prompt.

You want to structure it, keep it queryable, and feed in only what’s relevant.

2.4 Quality, Logs, and Evaluation

Day 4’s focus on quality and observability was humbling.

My earlier mindset:

“If it works on my test case, it’s probably fine.”

The course pushed me to:

• Log decisions and tool calls
• Trace multi-step workflows
• Treat evaluation as part of the design, not an afterthought

For my capstone, that meant:

• Looking at each agent’s input/output
• Detecting where hallucinations or inconsistencies appeared
• Thinking about simple checks, like:
  • Does the plan include all required sections?
  • Are financial assumptions at least internally consistent?

That mindset will matter a lot when I try to move from prototype to something SMEs can reliably use.

---

3. Surviving as a Non-Native English Learner

Important confession: English is not my working language.

That matters, because:

• The whitepapers are dense
• The podcasts are technical
• The notebooks are explained in English

If I had tried to brute-force everything, I probably would’ve dropped out by Day 2.

Instead, I built my own AI-powered learning stack on top of the course.

3.1 NotebookLM for Digesting Whitepapers

For each whitepaper, I:

1. Uploaded the PDF into NotebookLM
2. Asked it to generate a 5–7 minute Japanese explainer
3. Read the explainer to understand the structure and key points
4. Went back to the original text only when I needed details

This turned the whitepapers from:

“Walls of English text I must conquer”

into:

“References I can navigate with a guide”

3.2 Gemini as Code Tutor and Architect

When I hit code I didn’t understand, I pasted pieces into Gemini and asked questions like:

• “Explain this class as part of a workflow for SME business planning.”
• “What responsibility does this function have in the agent architecture?”
• “Is there a cleaner way to structure this multi-agent flow?”

Gemini helped me bridge:

• Low-level code understanding, and
• High-level architecture and business context

That bridge is what made it possible for me to attempt a serious capstone at all.

---

4. Capstone: AI STRATEGIC PLANNER for SME Business Plans

For the capstone project, I built:

AI STRATEGIC PLANNER

A multi-agent system that helps SME owners turn their story and website into a structured business plan.

Capstone write-up:

https://www.kaggle.com/competitions/agents-intensive-capstone-project/writeups/ai-strategic-planner-3357042

The core idea can be summarized as:

Your Story + URL + AI Agents = Your Own BMC

4.1 Problem: Business Plans Are a Cognitive Tax

In my daily work with SMEs, I see a recurring pattern:

• Founders have strong intuition about:
  • Customers
  • Products and services
  • Local competition
• But when they face:
  • Subsidy applications
  • “3-year revenue projections”
  • Business model templates

…their energy drops.

The format becomes a barrier.

Advisors like me help, but we’re not perfect either:

• Some are good at market analysis
• Some at finance
• Some at storytelling

Almost no one is great at all of it.

This is exactly the kind of “multi-perspective” problem that a team of agents can help with.

4.2 Design: Four Agents, Four Roles

Inspired by the course, I refused to build “one giant agent”.

Instead, I designed a team of four:

• Analyst

  • Reads the founder’s description and website
  • Extracts key facts: customers, value, strengths, weaknesses

• Strategist

  • Proposes a business model structure
  • Drafts a first Business Model Canvas (BMC)

• CFO

  • Creates rough financial assumptions
  • Checks basic feasibility: “Is this likely to be in the red or black?”

• Chief Editor

  • Takes everything above and produces:
  • A narrative business plan
  • Sections aligned with common subsidy/grant templates

Each role mirrors someone I’ve seen in real SME support:

• Analyst → junior consultant
• Strategist → business designer
• CFO → finance specialist
• Editor → the person who knows “how this needs to read on paper”

The course gave me the confidence to formalize these roles as actual agents, not just bullet points in my notebook.

4.3 Under the Hood: What I Used

At a high level, the system uses:

• Google’s Agent Development Kit (ADK)

  • To define and orchestrate agents

• MCP (Model Context Protocol)

  • To connect agents with external tools like note stores

• A simple memory layer

  • So later agents can read outputs from earlier ones

• Role-specific prompts

  • Heavily inspired by the “context engineering” and “quality” days

For every design decision, I asked:

“If this were a human team,

who would own this step,

and what would they need to see?”

That question came directly from how the course framed agents as structured workflows, not magic boxes.

---

5. How My Understanding of Agents Evolved

Looking back, my mental model went through four stages:

1. Before the course

   • “Agents = LLMs with bigger prompts + maybe some tools”

2. After Day 2 (Tools & MCP)

   • “Agents = LLMs that can call external systems in a structured way”

3. After Day 3–4 (Context & Quality)

   • “Agents = long-running workflows with memory, tools, and observable behavior”

4. After the Capstone

   • “Agents = teams of specialized workers, where architecture and evaluation matter as much as the model itself”

I now think less in terms of:

• “What prompt should I write?”

and more in terms of:

• “What roles exist in this problem domain?”
• “What tools do they need?”
• “What should they remember or forget?”
• “How will I know the system is behaving well?”

For me, that is a big upgrade compared to where I started.

---

6. What I’ll Take Back to My Day Job

The Intensive was not just an academic exercise.

I’m already thinking about how to adapt these ideas for SME support.

Some directions:

• Interactive business plan drafting

  • Founders talk; agents listen, structure, and draft a plan

• Scenario exploration

  • CFO-like agents quickly sketch “what-if” revenue and cost scenarios

• Knowledge reuse

  • Successful patterns encoded as starting points for new founders in similar industries

Most importantly:

I no longer see “agents” as futuristic or out of reach.

I see them as next-step tools to make expert support more accessible to the small businesses that need it most.

---

7. Final Reflection

If I had to summarize my experience in one line, it would be:

This course turned “agents” from a buzzword

into a mental toolbox I can actually use.

I joined the Intensive as:

• A non-native English speaker
• Not a full-time developer
• Someone who spends more time with business owners than with code

And I finished with:

• A working multi-agent prototype
• A clearer sense of agent architecture and evaluation
• A concrete plan for how agents can support real SME business planning

Huge thanks to the teams at Google and Kaggle for creating content that is both ambitious and practical.

And to anyone reading this who feels that agents are “too advanced” for them:

If an SME advisor from regional Japan can finish this course

and ship a multi-agent capstone,

you probably can, too.