Intro — A Course That Changed How I See Agentic Systems

When I started the 5-Day AI Agents Intensive with Google &amp; Kaggle, I expected another technical bootcamp.

I was wrong.

<a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F95g3c6xgkfbkjapd7wzy.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F95g3c6xgkfbkjapd7wzy.png" alt=" " width="800" height="1200"></a> 
The course wasn’t just “how to build an agent.” 
It reframed what an agent is — 
not a single script, not a chatbot, not a workflow…

…but a living architectural pattern:

perceiving

reasoning

deciding

acting

reflecting

and adapting over time.

By Day 5, something clicked for me:

“If agents fundamentally revolve around perception → memory → reasoning → action → feedback… then the tool we choose doesn’t matter as much as the architecture we build around it.”

That insight pushed me to design something more ambitious than just a capstone.

I built an entire agentic operating system using nothing but:

Google Workspace

Gemini (conceptually, until I can access paid API)

Apps Script

Node.js

and six GitHub repos working together.

This became my capstone — 
and honestly one of my favorite projects I’ve ever built.

<ol>
<li>What Resonated Most: Agents Are Not Single Agents — They Are SYSTEMS</li>
</ol>

Throughout the course, Google hammered several core topics:

1 — Event-driven architecture

Agents shouldn't always “run.” 
They should wake up when something meaningful happens.

2 — Memory as a first-class citizen

An agent without memory is a toy. 
An agent with structured memory becomes a system.

3 — Tools as capabilities, not dependencies

Agents should be able to pick the right tool at the right time.

4 — Multi-agent ecosystems

The real power is not a single agent, 
but multiple components interacting over shared state.

And THAT was the turning point for me.

Because I realized:

<a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F4eqg6vb8946vv4a7os69.webp" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F4eqg6vb8946vv4a7os69.webp" alt=" " width="800" height="1200"></a> 
Google Workspace is already a multi-agent ecosystem.

Gmail = event stream 
Docs = long-form memory 
Sheets = structured memory 
Forms = intake interface 
Apps Script = automation spine 
Gemini = reasoning engine

So instead of building one agent…

…I built a Google-native agentic OS.

<ol>
<li>My Capstone Project — MindsEye OS Across 6 Repos</li>
</ol>

Here is the full architecture (all repos are public):

<ol>
<li>Workspace Automation Layer</li>
</ol>

🔗 <a href="https://github.com/PEACEBINFLOW/mindseye-workspace-automation" rel="noopener noreferrer">https://github.com/PEACEBINFLOW/mindseye-workspace-automation</a>

Captures events from Gmail, Docs, Drive, Forms → normalizes → sends to ledger.

<ol>
<li>The Ledger (Prompt Evolution Tree + Runs)</li>
</ol>

🔗 <a href="https://github.com/PEACEBINFLOW/mindseye-google-ledger" rel="noopener noreferrer">https://github.com/PEACEBINFLOW/mindseye-google-ledger</a>

A time-labeled memory system (soft blockchain). 
Nodes = prompts / ideas. 
Runs = executions / results.

<ol>
<li>Gemini Orchestrator (Kernel)</li>
</ol>

🔗 <a href="https://github.com/PEACEBINFLOW/mindseye-gemini-orchestrator" rel="noopener noreferrer">https://github.com/PEACEBINFLOW/mindseye-gemini-orchestrator</a>

This is where Gemini will live as the reasoning engine.

<ol>
<li>Devlog Generator</li>
</ol>

🔗 <a href="https://github.com/PEACEBINFLOW/mindseye-google-devlog" rel="noopener noreferrer">https://github.com/PEACEBINFLOW/mindseye-google-devlog</a>

Turns the ledger into weekly narrative summaries.

<ol>
<li>Analytics Engine</li>
</ol>

🔗 <a href="https://github.com/PEACEBINFLOW/mindseye-google-analytics" rel="noopener noreferrer">https://github.com/PEACEBINFLOW/mindseye-google-analytics</a>

Exports → Looker dashboards → KPI insights.

<ol>
<li>Workflow Atlas (Portal Maps)</li>
</ol>

🔗 <a href="https://github.com/PEACEBINFLOW/mindseye-google-workflows" rel="noopener noreferrer">https://github.com/PEACEBINFLOW/mindseye-google-workflows</a>

Everything described as YAML “portal flows” so the whole OS can be understood.

<ol>
<li>How My Understanding of Agents Evolved
Before the course:</li>
</ol>

“Agents are programs that take an input and produce a dynamic output.”

After the course:

“Agents are architectures built around structured memory, perception streams, execution graphs, and feedback loops.”

<a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F1qyz1srx2bvnev4sp7zw.webp" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F1qyz1srx2bvnev4sp7zw.webp" alt=" " width="800" height="1200"></a> 
This shift shaped my design:

<ol>
<li>Mapping Course Concepts → My MindsEye OS</li>
</ol>

Here’s how each idea in the course shaped each repo.

4.1 Perception: Workspace Automation Repo

Course Concept: 
Agents start with perception channels — email triggers, file events, user actions.

My Implementation: 
I built an event listener layer using Apps Script:

Gmail label “Send to MindsEye”

Docs custom menu

Drive folder watchers

Form submissions

Every event becomes:

{ 
 "surface": "gmail", 
 "event_type": "label_applied", 
 "title": "...", 
 "source_url": "...", 
 "summary": "...", 
 "timestamp": "..." 
}

This normalized event is the perception signal.

4.2 Memory: The Ledger Repo

Course Concept: 
Agents evolve when their memory evolves. 
Memory = core.

My Implementation: 
A Google Sheet (nodes + runs) functioning as a soft blockchain:

Nodes = ideas, prompts, branches

Runs = experiments, AI outputs, score, context

Everything is append-only + time-labeled.

The result:

A living, multi-user, multi-team evolutionary tree of intelligence.

4.3 Reasoning: The Gemini Orchestrator Repo

Course Concept: 
Reasoning = choosing a strategy from memory + context.

My Implementation: 
mindseye-gemini-orchestrator is a Node.js kernel:

Selects nodes ready to run

Passes metadata + linked Docs into Gemini

Logs run results back as new memories

Once I have Gemini credits, it becomes an actual functioning kernel.

4.4 Reflection: Devlog Repo

Course Concept: 
Agents should reflect, explain, summarize.

My Implementation: 
The Devlog generator groups runs, detects patterns, and writes weekly summaries.

<a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fxzyddm53jmkr4a46qqet.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fxzyddm53jmkr4a46qqet.png" alt=" " width="800" height="1200"></a> 
It turns raw memory into narrative memory.

4.5 Insight: Analytics Repo

Course Concept: 
Agents need feedback loops.

My Implementation: 
Exports → Python scripts → Looker dashboards:

run frequency

success rates

node lineage

surface usage

prompt performance

This becomes the intelligence lens across 200+ users.

4.6 Orchestration: Workflows Repo

Course Concept: 
Multi-agent systems require explicit coordination, NOT magic.

My Implementation: 
I built YAML “portal maps” describing flows like:

trigger: workspace_event 
steps:

<ul>
<li>normalize_event</li>
<li>write_to_nodes</li>
<li>schedule_gemini_run</li>
<li>generate_devlog</li>
<li>refresh_analytics</li>
</ul>

This repo is the architecture documentation for MindsEye OS.

<a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fe0dqripmapbi672cdsph.webp" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fe0dqripmapbi672cdsph.webp" alt=" " width="800" height="1200"></a>

<ol>
<li>What I Learned (In One Sentence)</li>
</ol>

The future of AI isn’t one agent — it’s an ecosystem of interconnected subsystems that perceive, remember, reason, act, and reflect in coordinated loops.

This course gave me the mental model. 
My six repos gave me the playground.

<ol>
<li>Why This Matters (Especially for Teams of 200)</li>
</ol>

I designed the system so:

200 people can create their own prompt trees

their data never overwrites each other

all nodes and runs become shared, structured memory

patterns flow between teams even when raw data is siloed

Gemini becomes the reasoning core for the whole organization

It becomes a Google-native AI operating system.

Lightweight. 
Auditable. 
Extensible. 
Future-proof. 
And once Gemini API access becomes available — fully agentic.

<ol>
<li>Final Reflection — “I Didn’t Build an Agent. I Built an OS.”</li>
</ol>

The course didn’t just teach me how to build agents.

It taught me how to think like a system designer.

It pushed me to:

move from 1 agent → multi-agent

move from scripts → architecture

move from outputs → memory

move from tools → capabilities

move from workflows → operating systems

This became my capstone.

Six repos. 
One architecture. 
A Google-native MindsEye OS. 
Built with no cloud budget, no paid credits — just design and pattern recognition.

And I’m only getting started.

<a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fwho3w4gudncyv0lxyuv8.webp" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fwho3w4gudncyv0lxyuv8.webp" alt=" " width="800" height="1200"></a>

تتناول المقالة تجربة تحويلية خلال دورة مكثفة لمدة 5 أيام حول وكلاء الذكاء الاصطناعي التي نظمتها جوجل وكاجل. في البداية، توقع الكاتب أن تكون دورة تقنية، لكنه اكتشف فهمًا أعمق للأنظمة الوكيلة. أكدت الدورة أن الوكيل ليس مجرد نص أو روبوت محادثة، بل هو نمط معماري ديناميكي يتضمن الإدراك، والتفكير، واتخاذ القرار، والعمل، والتفكير العميق، والتكيف مع مرور الوقت. كان لهذا الإطار الجديد لما يشكله الوكيل تأثير كبير على وجهة نظر الكاتب حول تطوير الذكاء الاصطناعي.

El artículo narra una experiencia transformadora durante el curso intensivo de 5 días sobre agentes de IA organizado por Google y Kaggle. Inicialmente, el autor esperaba un bootcamp técnico, pero descubrió una comprensión más profunda de los sistemas agentivos. El curso enfatizó que un agente no es simplemente un script o un chatbot, sino un patrón arquitectónico dinámico que involucra percepción, razonamiento, toma de decisiones, acción, reflexión y adaptación a lo largo del tiempo. Este nuevo enfoque sobre lo que constituye un agente impactó significativamente la perspectiva del autor sobre el desarrollo de la IA.

L'article raconte une expérience transformative lors du cours intensif de 5 jours sur les agents IA organisé par Google et Kaggle. Au départ, l'auteur s'attendait à un bootcamp technique, mais il a découvert une compréhension plus profonde des systèmes agentiques. Le cours a souligné qu'un agent n'est pas simplement un script ou un chatbot, mais un modèle architectural dynamique impliquant la perception, le raisonnement, la prise de décision, l'action, la réflexion et l'adaptation au fil du temps. Ce nouveau cadre de ce qui constitue un agent a eu un impact significatif sur la perspective de l'auteur sur le développement de l'IA.

The article recounts a transformative experience during the 5-Day AI Agents Intensive course organized by Google and Kaggle. Initially expecting a technical bootcamp, the author discovered a deeper understanding of agentic systems. The course emphasized that an agent is not merely a script or chatbot but a dynamic architectural pattern involving perception, reasoning, decision-making, action, reflection, and adaptation over time. This reframing of what constitutes an agent significantly impacted the author's perspective on AI development.

My 5-Day AI Agents Intensive Journey — And How I Built a Google-Native MindsEye OS Using 6 Repos

Was this article worth reading? Share it

Ready to build your own newsroom?