Blog · 5 min read ·
Case Study: Building a Platform With AI Agents, Solo
Direct answer: Akkademi is an online question-practice platform for the ALES and DGS exams in Türkiye. One person built it, without a software team, using AI agents. It has a website, a mobile app, memberships and payments, legal texts and ad measurement.
Can one person build a product with AI agents?
Yes, but not by pressing a button. The founder made the product decisions, coordinated the work, reviewed it and approved every release. The agents did the repetitive execution: writing code, preparing releases, drafting designs, texts and measurement set-ups.
We do not publish user numbers, revenue or customer data for Akkademi, and we make no promise that you would get the same result. This article is about the method. You can find the project on our work page.
How the work was organised
Agents need boundaries. The work was split into four areas, each with its own agent:
- Code
- Release and server
- Design and content
- Text and measurement
All agents read one shared rules file. It was short:
- No passwords written anywhere.
- Never touch another agent's folder.
- Ask before deleting anything.
- Write a dated report folder for every job.
The dated reports gave the founder a trail to follow. Anthropic's Building effective agents describes a useful distinction. Workflows are systems where LLMs and tools follow predefined code paths. Agents are systems where the LLM directs its own process and tool use. Risky steps, like releases, followed a fixed path. Looser agent work was kept for drafts.
Why we write this down
Most stories about AI-built software skip the dull parts. We think the dull parts are the useful ones: who did what, which rules the agents followed, what went wrong and how a release was kept safe. If you plan to work with agents yourself, these are the points that decide whether it stays under control.
The release gate
Nothing went live without passing the same sequence, every time.
| Step | What happens | Human role |
|---|---|---|
| Backup | A copy is taken before any change | Check that it exists |
| Tests and gate | Automated tests and the release checks run | Read the result |
| Release | The change goes live | Approve |
| Live check | The live site is tested again | Check the result |
| Rollback | The old version returns if the live check fails | Decide |
One person cannot rely on memory under pressure. The gate was written down so that nobody had to remember it.
Legal and privacy texts
The legal texts were drafted with the agents. Drafting was only the first step. The founder then compared each text line by line with how the site actually behaves: what the forms collect, when measurement runs and what users can ask for. The final decision on every legal text went to a lawyer.
The platform has no fake reviews, no fake counters and no fake discounts. A legal page has to describe what the site does, not the image the owner would like.
Mistakes happened
Agents do not remove mistakes. Three kinds came up:
- Colliding patches. Two agents changed overlapping things, and the changes clashed.
- A test timeout. A test ran out of time and blocked a release until it was understood.
- A misread task. An agent understood a request differently from what the founder meant, and the work had to be redone.
Each of these led to a change in the rules. That is why the rules file and the gate are not paperwork.
Notice what is missing from this story: a claim that the agents worked flawlessly. They did not, and the process was built on that assumption.
Four lessons
- Separate the folders. Agents are safer when they cannot get in each other's way.
- Write the gate before the release. Backup, tests, live check and rollback should be ready before there is any pressure.
- A draft is not an approval. Legal, privacy and payment wording needs a person, and for legal texts a lawyer.
- When in doubt, roll back. Returning to the last working version is faster than arguing over a log.
AI does not do everything on its own here. The founder still reads reports, answers questions the agents cannot settle and says no when a change is not ready. Take that person away and the setup stops working. It speeds up execution, and a person stays responsible.
Using the same method on your site
We use the same structure for existing websites and apps. You do not need a new platform to benefit from it:
- Audit. A passive, permitted check of your site, with scores and a priority list. Start with the audit report, from $490.
- Fix. Small repairs follow the same rules: a written list, a backup and a way back.
- Care. Monthly care keeps the site from drifting, with regular checks.
How we work describes the steps in detail.
FAQ
Can a solo founder really build a product with AI agents?
With clear rules, yes. Akkademi shows one person building a platform with a website, an app, payments and legal texts. The agents did execution, and the founder made decisions and approved releases.
Does AI do everything by itself?
No. The agents worked from a shared rules file, wrote dated reports and passed a release gate that a person approved. Legal texts went to a lawyer.
What was the release process?
Backup, tests and gate, release, live check and rollback if the check failed.
Where does the lawyer fit in?
The agents drafted. The founder compared each draft with the real behaviour of the site. The lawyer made the final call.
Does this work for an existing website?
Yes. We apply the same method to audits, fixes and monthly care.