From Hello World to a fleet of AI agents: what changed in two years

I am not a developer. Two years ago, I shipped an API in a couple of days. Today AI agents build it - and check each other.

Background

I never had a passion for coding. The only time I really wrote code was when I typed “Hello World” and immediately understood it was not for me. Early in my career I did have to modify scripts - Bash, VBScript, PowerShell, and specialized ones like Veritas NetBackup and Oracle RMAN - but that was editing, not building.

Discovery

I first started using OpenAI’s models before they were widely known, through the Microsoft for Startups Founders Hub program. When ChatGPT launched, its friendly interface made the model far easier to use than working with the OpenAI API directly. Then in 2023 OpenAI announced custom GPTs, which let you combine a model with your own knowledge. Working on a hobby project, I hit a wall: I needed current economic and market data that the model simply could not see. Connecting an external API to the GPT looked like the answer. The market for that data is broad, from cheap basic feeds to mid-tier aggregators to premium direct-exchange feeds.

Beginning

So, I wondered: what if I just asked ChatGPT whether it could help me build the API itself? I opened ChatGPT Plus, asked a simple question, and got a confident answer: yes, of course, let us get started. That began my adventure into service design, functionality, and logic. With four tools - ChatGPT Plus, GitHub Copilot, Visual Studio Code, and, now and then, my own brain - I built a working API that pulled the live data I needed. Deploying it was easy: it ran on Azure App Service on Linux, with minimal effort and cost. The 2024 prototype was around five thousand lines, and a standard code review came back with good marks and only minor formatting notes.

The result: I published the Finance Counselor GPT in the public GPT Store, with a working service behind it at api.sugra.ai

The point

If a newcomer far from programming could launch a real cloud service that fast, imagine what a career developer can do. AI does not replace the professional - it makes the work faster, better, and lighter.

Two years on, that same instinct compounded into something I did not expect, and it quietly grew into Sugra API - no longer a finance tool, but domain-agnostic intelligence infrastructure. Finance turned out to be just one field among many, sitting alongside macro, entities, networks, news, earth, and research, all of it drawing on primary and sovereign sources: central banks, regulators, national statistical agencies, intergovernmental bodies, and universities. I did not set out to build that. It grew.

But the deeper change is in how it gets built. These days the shape of the work is simple to describe: one person making the decisions, and a fleet of agents doing the rest. Three threads tell that story.

Thread one: I no longer type the code

I direct several autonomous AI agents working in parallel, each in its own isolated branch so they never collide. Every change rides the same line - claim, build, review, verify, ship. My job is intent, sequencing, and judgment. Theirs is implementation.

Thread two: quality is no longer one review pass

It is guarded by cross-vendor review gates. The work is checked by AI agents from three competing companies - Anthropic, OpenAI, and Google - chosen because they were trained differently and carry different blind spots. The reviewer is, by design, a competitor’s model, so when one is wrong a genuinely different one catches it. Each change gets a second opinion before any code is written, tests before the fix, and rounds of review where every finding has to be closed before it can merge. This is not theory: concurrent cross-vendor review has already caught a real shipped data bug, and a wrong assumption about a market’s closing time that a single review had already approved. A green pipeline only proves the code left; telemetry proves it arrived, and nothing is called shipped until it is verified live.

Thread three: the simple GPT app became a product

Sugra API (sugra.ai) is now running in three customer pilots - real teams testing it against decisions that matter. I will save the specifics for when they are earned.

And this is only my small corner of it. Imagine a world where even routine, repetitive work is handled well by machines - from medicine, where AI helps catch disease early, to education, to the daily work of managing our own time and attention.

A few years ago, the Code.org initiative, backed by everyone from Mark Zuckerberg to Bill Gates to Barack Obama, urged the world to learn coding. Many parents wanted their children to become coders, believing the future needed people who could write code. The goal was inspiring, but it centered on coding rather than on analysis, judgment, and clear thinking. More recently, Nvidia CEO Jensen Huang argued almost the opposite at the World Governments Summit in Dubai: kids should not need to learn coding, let AI do it, and free people to focus on biology, education, manufacturing, economics. The future is already here, it is just not evenly distributed, as William Gibson put it.

And yet, a point Sam Altman has made publicly stuck with me: do not underestimate the human element. Do not bet on AI alone, and do not expect more from it than it can give. He is right, and watching my own work change has only convinced me more. The human did not disappear when I stopped typing code - it moved up a level, from writing lines to directing a fleet, from execution to judgment. I still own the intent, the approvals, every change that touches production, and the calls that have no clean answer. The deciding stays human.

So, the qualities we tend to underrate - humanity, empathy, the ability to adapt under stress, to think clearly and say what you mean - do not fade as the machines get better. They become the job. We can have knowledge, technical skill, and the power to build supercomputers, and still fall short on exactly these. As the coder becomes the director of agents, those qualities matter more, not less. AI will not supply them for us.

Thinking about the future means thinking about the present - looking at your life differently, and reasoning on a different scale.

Arman Obosyan

Originally published on LinkedIn, 14 April 2024.