The dream of “autonomous coding” is closer than ever, but as any seasoned engineer knows, the devil is in the deployment.
Recently, I decided to put some of the most advanced AI tools to the test Antigravity Pro (using Claude 4.5 Thinking/Opus) and Gemini 3 Pro. I wanted to see if an AI agent could navigate a complex, real-world SaaS environment without hand-holding.
The results were impressive, chaotic, and a massive wake-up call for the future of our profession.
The Setup: A Multi-Headed SaaS Beast
The project wasn’t a simple “To-Do” app. It’s a complex SaaS ecosystem featuring:
One Backend powering five different Frontend systems (Project, Compliance, Finance Management, etc.).
AWS CDK Infrastructure: Meaning the backend cannot be run locally; it must be deployed to the cloud for testing.
Strict Workflows: Technical docs must be approved by Seniors, PMs, and DevOps before a single line of code is written.
I gave the AI a User Story for a large feature in the Compliance Management module and a very detailed prompt. However, I intentionally withheld four “human” insights about our specific environment:
The inability to test locally.
Deployment is only allowed from the
dev-tempbranch.The Service Stack file was at its limit (needing a new stack).
A “Double-Deploy” requirement to prevent CORS issues with AWS Cognito.
The Good: Coding Standards and UI Precision
The AI started strong. It analyzed the database schema, updated the scripts, and built the backend following our existing patterns (Dependency Injection, RESTful APIs, etc.).
When it moved to the frontend, I was genuinely impressed. Using only Figma screenshots, it built a beautiful, responsive UI that matched our design system perfectly. It refactored mock data into real API calls using Axios without breaking a sweat. 48 files were modified or created, and the code quality was top-tier.
The Bad: When the “Agent” Lost the Plot
The friction started during the deployment phase. This is where the difference between a coder and an engineer became clear:
Branch Blindness: It tried to deploy from my local
dev-lakshanbranch. After five minutes of "thinking" through errors, it correctly identified that it needed to switch todev-temp.The Resource Limit: It hit the AWS Service Stack limit. To its credit, it eventually figured out it needed to create a new stack.
The CORS Nightmare: Because it didn’t know about the “double-deploy” rule for Cognito, it hit CORS errors during testing. Instead of questioning the infrastructure, the AI went down a rabbit hole altering pre-configuration files and trying to “bypass” AWS security.
The Warning Sign: If I had blindly approved those AWS changes, the entire project infrastructure would have been compromised. The AI began “hallucinating” infrastructure fixes that were actually destructive.
The “Human” Fix
I eventually had to intervene. I updated the prompt: “Once everything is complete, deploy the backend twice.” Suddenly, everything worked. Even the Pull Request (PR) process needed a human touch; the AI initially tried to merge dev-temp directly into the main/dev branch instead of returning to the feature branch first.
Key Takeaways for the AI Era
UI is AI’s Strong Suit: Frontend development and design-to-code translations are becoming incredibly efficient.
Infrastructure is the Final Frontier: AI still struggles with “tribal knowledge” those specific quirks of a company’s AWS/DevOps setup that aren’t written in the documentation.
The “Senior” Filter: You cannot give these tools to someone who doesn’t understand the foundation. An inexperienced dev would have pushed the “CORS bypass” and broken the production environment.
Final Thoughts: Will SE Jobs Vanish?
No, but they are changing. The “Coder” who just translates requirements to syntax is at risk. However, the Software Engineer the one with the Basic + Foundation + Skills + Experience will become a “Pilot.”
AI will do the heavy lifting, but we are the ones who know when the plane is flying into a storm. To survive the AI wave, don’t just learn to code; learn how the whole machine works.
