This is a submission for the Google AI Agents Writing Challenge: Learning Reflections
Building a Real Business with Agentic AI
Advancements in Artificial Intelligence have quietly - and sometimes dramatically - redrawn the boundaries of human creativity 🎨. Domains once constrained by time, effort, and human bandwidth, such as - digital art, music, poetry, and design - have expanded at an unprecedented pace. What felt impossible yesterday is now not only achievable, but repeatable.
As the founder of a company focused on customised and deeply personalised digital gifts, I found myself standing at the intersection of creativity and automation. Our belief was simple, yet ambitious:
If AI can generate intelligence at scale, why can’t it make every customer feel individually understood at scale?
What followed was not just a technical experiment, but a formative learning experience - one that reshaped how I think about systems, scale, reliability, and human trust.
🌱 The Experiment That Sparked Everything
At Invysia, personalisation is not a feature layered on top of a product—it is the product. Every calendar, every digital artwork, every personalised website is created uniquely for the customer. No templates reused. No shortcuts taken.
Inspired by the rapid advances in AI-driven creativity, we launched a pilot project: invysia.store. The vision was bold:
Use AI to create and deliver highly personalised digital products end-to-end, with minimal human intervention.
On the surface, it worked:
- Calendars were generated from scratch
- Designs were tailored to individual stories
- Customers loved the uniqueness
But operationally, cracks began to appear.
Behind the scenes, we were still:
- Spending hours conversing with customers to understand intent
- Manually refining prompts
- Relaying feedback between AI outputs and user expectations
- Acting as an invisible orchestration layer
The uncomfortable truth became clear:
AI had accelerated creation - but humans were still coordinating everything.
This was not scalable. And more importantly, it was fragile.
🧠 The Realisation: Creativity Scales, Coordination Does Not
Despite my background in computer science, agentic AI initially felt abstract and intimidating. There were too many choices - models, frameworks, orchestration styles - each promising autonomy, yet offering little clarity on control, reliability, and cost.
The idea of running a business end-to-end using autonomous AI agents raised hard questions:
- How do you trust non-deterministic systems?
- How do you debug decisions made by probabilistic models?
- How do you prevent runaway costs?
- How do you ensure continuity when an agent fails?
These questions stalled progress - until I discovered the Google 5-Day AI Agents Intensive on Kaggle.
🎯 The Turning Point: Learning to Engineer Agents, Not Just Prompt Them
This course did not glorify agents as magical entities. Instead, it treated them as production systems.
Over five days, agentic AI was broken down into engineering primitives:
- Models
- Tools
- Orchestration
- Context & Memory
- Evaluation
- Deployment
What surprised me most was that nearly 75% of the course focused on what traditional software engineers already care deeply about:
- Observability (logs, traces, metrics)
- Reliability and evaluation
- Cost optimisation
- Security and governance
- Production readiness
That was the missing piece. Agentic AI was not abandoning software engineering - it was extending it.
🤖 Applying the Learnings: Building a Real Multi-Agent Business System
Armed with this clarity, I rebuilt the core of Invysia as a multi-agent system, not a single all-knowing chatbot.
Two agents. Clear boundaries. One coherent system.
🟣 Iris — Sales & Support Agent
Iris is the first point of contact. She:
- Handles customer inquiries and FAQs
- Explains pricing and timelines
- Conducts requirement discovery through structured questions
Crucially, Iris does not pass raw chat logs downstream.
Instead, she performs context engineering.
From unstructured conversations, Iris extracts structured intent - a clean, validated JSON object containing:
- Theme preferences
- Color palettes
- Aspect ratio and resolution
- Delivery timelines
This structured context is stored in ADK Session State, allowing it to persist reliably across agent boundaries. The outcome is not “shared memory” in the abstract - it is shared understanding.
🟠 Daedalus — Designer Agent
Daedalus never talks to the customer. He receives only structured intent from Iris.
His responsibilities include:
- Generating creative prompts using a dedicated prompt-generator sub-agent
- Producing 12 monthly calendar designs asynchronously
- Handling payment link generation and delivery
Importantly, Daedalus is not a free-form creative chatbot.
He is strictly bound by tools:
- A specific Image Generation Tool
- A Template System that enforces layout, typography, and aspect ratio
This ensures Daedalus behaves like a software operator, not a hallucinating artist. He cannot invent formats, bypass templates, or deviate from constraints, but still creates engaging artworks.
🧩 Learning 1: Designing Boundaries Matters More Than Designing Intelligence
One of the most important lessons I learned is that agent reliability comes from boundaries, not brilliance.
Early on, I tried to make agents smarter—better prompts, deeper reasoning, larger models. Most failures, however, came from unclear responsibility, not lack of intelligence.
So I flipped the question:
“What is this agent not allowed to do?”
- Iris never designs
- Daedalus never sells
- Each agent has permissioned tools and scoped authority
This reduced hallucinations, simplified debugging, and made behavior predictable.
🧠 Key insight: I stopped trying to make agents smarter and started making them more disciplined.
⚡ Learning 2: Speed Comes from Parallelism, Not Better Models
Another breakthrough was realising that latency is an architectural problem.
Initially, workflows were linear:
- Prompt creation blocked image generation
- Each step waited on the previous one
By redesigning the system:
- Prompt generation runs in parallel
- Image creation happens asynchronously
- Agent handoffs are non-blocking
The result was dramatic.
What once took 3–5 days now completes in under 5 minutes - without upgrading models or increasing cost.
⚡ Key insight: The biggest performance gains came from letting agents work in parallel - just like effective human teams.
🛠️ Learning 3: Cost Optimisation Is a Design Choice, Not a Post-Processing Step
The course emphasised that model selection is architecture, not configuration.
I applied this directly:
- Gemini 2.5 Flash powers Iris - fast, lightweight, cost-efficient for frequent conversations
- Gemini 3 Pro Image powers Daedalus - used only after conversion, where high-quality output matters
Heavy compute is triggered only when value is guaranteed.
💰 Key insight: Every agent decision has an economic footprint.
🧍 Learning 4: Human Fallback Is a Reliability Feature, Not a Weakness
Autonomy without trust slows adoption.
The system was designed with human fallback mechanisms:
- If Iris fails, a human can step in without breaking the flow
- If Daedalus fails, a human designer can take over using the same structured context
The pipeline never collapses-it degrades gracefully.
This mirrors enterprise reliability patterns and builds confidence for both internal teams and customers.
🤝 Key insight: Human-in-the-loop is not a crutch; it is a trust accelerator.
🌍 How These Learnings Benefit Everyone Involved
Customers
- Near-instant delivery
- Consistent quality
- Truly personalised designs
Team
- Reduced operational fatigue
- Clear intervention points
- Focus on creativity, not coordination
Builders & Founders
- Proof that agentic AI is not hype - it is engineering
- A blueprint for scaling creativity responsibly
- Confidence that small teams can build production-grade systems
✨ Closing Reflection
This journey taught me that the real power of agentic AI is not automation - it is alignment.
Alignment between:
- Intent and execution
- Creativity and control
- Human values and machine efficiency
Top comments (0)