An experiment in bridging the gap between concept and code: building a non-trivial system using Agent-assisted development.
Having returned to South Africa permanently after almost 20 years in the UK, I found myself with some time to kill whilst staying at my mother’s house. With all my worldly possessions on a boat somewhere off the coast of Liberia making their way to me, I commandeered my mother’s laptop and started having a serious look at Google’s new "IDE," Anti Gravity.
Having previously used Gemini to produce a REST API specification to get around a resourcing constraint at KPMG—which amazingly, made it into production—I was keen to see what could be achieved at a systems level. The REST APIs and related examples at KPMG were small, isolated tasks; I was interested in what could be reasonably achieved when using GenAI to produce a non-trivial system.
My first choice was the language and related frameworks. The language I started my career developing in has been functionally dead for years, so I narrowed it down to JavaScript (TypeScript) or Python. Both are very well-supported languages with large communities, endless libraries, and many examples on Stack Overflow—so good for both my capability and the LLM's.
When there is a large intake at a warehouse, there are countless mismatches between the physical item (including tags and labels) and the documentation. This drastically reduces how quickly the stock consignment can be receipted and ultimately made available for sale.
If only there was something that could take a product identifier, like a SKU, barcode, or description, and return the data you need, collected from all publicly available information about the product on the web.
I designed a self-administered system that allows you to create API endpoints and define output parameters (like colour code, material, and dimensions). Once your API is set up and a request is sent, your submitted descriptor is passed to Gemini via a detailed prompt. Gemini runs the prompt, collects all available information, intelligently maps it to your specified output parameters, and returns it in the response JSON.
To have a reasonable chance of success, I wanted to create a small application, but one with a non-trivial amount of complexity. The system features Clerk Authentication, a Postgres DB on Supabase, Google Gemini integration, and management of organisations, APIs, routes, and endpoints.
Anti Gravity is very new, so online resources are limited to simple 'one-shot' tasks. I used the "echo technique" to get started, asking Google Gemini what the best approach was and chatting until we had a PRD-type document. This gave a good high-level view of what we intended to achieve, including the tech stack decisions.
I used a similar process for the prompt, using the echo technique and the saved PRD from the previous chat to build it out.
I decided to go with Agent-assisted development. I didn't go with the full Google-defined system involving an orchestrator and sub-agents; that arrangement seemed a step too far for a first try. Instead, I loaded the PRD into Anti Gravity, which included the DB creation scripts and code examples I had put together with Gemini.
After a few minutes and many tokens, Anti Gravity produced an implementation plan. We didn't follow it exactly. Drawing on my experience in architecture and product management, I slowly asked the agent to build out portions of the application in the sequence I would have used with a team of human developers:
By this point, I had probably spent about 2.5 days working on the application. Thinking of the median velocity of teams I have managed, I think this would have taken about a week with a team of 4-5 developers. So, yes, AI is going to change our jobs a lot in the coming decade.
The willingness of the AI to help solve your problems is so enthusiastic that it regularly copies your connection strings, passwords, keys, and secrets into test files which will get pushed to GitHub if you don't pay attention. This happened at least five times, so it is worth paying extra attention to this known problem.
The old adage that technical proficiency in a PM comes at the expense of core product skills is truly obsolete—and, in my view, was always a misconception. In the coming decade, PMs who genuinely understand the underlying technology can leverage AI to bridge the gap between concept and code.
We are moving toward a reality where a specification isn't just a document, but a working application; or, if the product is lean enough, the PM might build the whole thing themselves. This shift also challenges the notion that execution is secondary to strategy. Strategy and execution have always been symbiotic; to ensure a vision comes to life as intended, you have to be close to the build.
To succeed in this new era, being "end-to-end" isn't just an advantage—it’s a requirement.