🚀 Working as a Dev with 100% AI: My Real Experience Inside a Bank
If you had told me some time ago that I could work as a developer without actually developing, I would probably have laughed and moved on.
But today, not only is this possible, it’s already happening.
And I’m living this in practice.
🧩 The Context
I’m part of a squad in a bank here in Brazil that was chosen as a pilot to test a new way of working:
👉 Using AI as the main driver of development, not just an assistant.
No more using AI just to ask questions or generate a snippet here and there.
The idea was bold:
💡 What if AI did 100% of the technical work, and the developer became an orchestrator?
Spoiler: it works better than it sounds.
🤖 How I Used AI Before
My use of AI was pretty much what most devs do:
- Ask something
- Get an answer
- Copy what makes sense
- Adjust it manually
In other words:
AI was basically a boosted StackOverflow
Useful? A lot.
Transformational? Not really.
⚙️ The New Model: AI as the Executor
Everything changed when we started structuring how AI is used inside the development flow.
Two concepts made all the difference:
📄 AGENTS.md
Think of it as:
🧠 A behavior manual for the AI inside your repository
It defines:
- Project structure
- Coding standards
- Conventions
- Business rules
- How to run, test, and validate
👉 Basically, the AI’s onboarding document
🧠 SKILLS
This is where the real power is.
Skills are capabilities that allow the AI to execute complete tasks, such as:
- Creating endpoints
- Fixing bugs
- Refactoring code
- Writing tests
- Analyzing logs and identifying issues
The shift is simple, but powerful:
❌ “How do I do this?”
✅ “Solve this.”
🔄 My Role Completely Changed
Before, I used to code.
Now, I:
- Define the problem
- Provide context to the AI
- Validate the results
- Adjust direction when needed
I’ve become something between:
- 👨💻 Dev
- 🧠 Product thinker
- 🤖 AI orchestrator
And that changes everything.
📈 The Impact in Practice
🚀 Insane Speed
Things that used to take days now take hours.
No exaggeration.
AI can:
- Understand the project context
- Apply patterns correctly
- Implement complete solutions
👉 And the best part: no repetitive manual work
🧹 Goodbye Manual Work
You know those tasks that aren’t really development?
Like:
- Filling out detailed PBIs
- Writing documentation manually
- Creating boilerplate
- Adjusting repetitive code
👉 AI just absorbed all of that.
And honestly, that was one of the biggest wins for me.
🐞 Smarter Debugging
With well-defined context (AGENTS.md + SKILLS), AI can:
- Analyze errors
- Suggest fixes
- Apply fixes directly
It feels like:
Having a senior developer available all the time — just faster.
⚠️ It’s Not Perfect (Yet)
Of course, there are challenges:
- AI needs well-written context, otherwise it gets lost
- Sometimes it solves things, but not in the best way
- You need to validate, you can’t blindly trust it
Which leads to an important shift:
You stop coding, but you need to understand the system much more deeply
💡 What Surprised Me the Most
It wasn’t just the speed.
It was realizing that:
🚨 The bottleneck is no longer writing code
Now the real bottleneck is:
- Clarity of the problem
- Quality of the context
- Ability to guide the AI
And that completely changes the developer profile.
🔮 The Future (My Take)
After living this, I have a strong opinion:
- Developers who only code will lose space
- Developers who can think, structure, and guide AI will grow fast
👉 Code is no longer the end goal.
👉 It’s just the means.
🏁 Conclusion
This experience inside my squad was a turning point.
I went from someone who used AI as support to someone who:
💥 Works with AI as the main execution layer
And honestly?
I don’t see this going back.
👇 Final Thought
If you’re a developer, here’s my advice:
Stop asking:
“How do I do this?”
Start asking:
“Do this for me — here’s the context.”
Because the game has already changed.
Top comments (1)
Awesome, especially the shift from "AI as assistant" to "AI as executor".
I've seen similar gains when context is extremely well-structured, but I'm curious about the edge cases in your bank pilot:
When the AI “owns” implementation end-to-end, currently, we may have trouble handling:
Would love to hear how your squad draws that line in practice.🙂