Picture this: You want to build a gaming portal. You need to find the games, playtest them, categorize them, write descriptions, extract thumbnails, and publish them. It is a massive bottleneck.

When I set out to build minigames.world earlier this month, I didn’t want to build just another directory. I wanted to build a system that manages itself.

Instead of a traditional CRUD application where I do all the data entry, I built an agentic architecture where AI analyzes, curates, and manages the entire game portfolio autonomously. Here is a breakdown of how the architecture works, the tech stack I chose, and why AI agents are fundamentally changing system design.The Problem: AI as a Feature vs. AI as an OrchestratorMost applications treat AI as an API call—you send text, you get text back.

To build an autonomous platform, the AI cannot just be a feature; it needs to be the orchestrator. The agent needs to be able to fetch a game, analyze its mechanics, determine its genre, decide if it meets quality standards, and execute the database updates to publish it.

The Architecture: Context, Control, and Concurrency

Running autonomous agents requires a backend that can handle massive concurrency and fault tolerance. If an agent hangs while analyzing a heavy WebGL game, it shouldn't bring down the system.Because of this, I rely heavily on Erlang and the BEAM ecosystem. The actor model is uniquely suited for agentic workflows. Each AI agent can run as an isolated, lightweight process. If an agent fails to parse a game's metadata or times out on an API call, the supervisor simply restarts it without affecting the rest of the orchestration layer.

Here is the basic workflow the agents execute:

1. Discovery & Ingestion: A worker process identifies a new game from a source feed.
2. Agent Analysis: The core agent steps in. It evaluates the provided metadata, parses the game's assets, and structures the data.
3. Curation Gate: This is the control layer. The agent evaluates the game against a set of quality heuristics. If it passes, the agent categorizes it.
4. Publication: The agent commits the new entry to the database.

Serving it Fast: Leaving Heavy Frameworks Behind

While the backend handles the complex orchestration of AI agents, the frontend needs to be aggressively optimized. Gamers want to click and play immediately.

I bypassed heavy client-side JavaScript frameworks. The frontend is built for pure performance, serving pre-rendered static structures with Astro and managing dynamic client interactions with HTMX. This keeps the time-to-interactive incredibly low and ensures the platform's SEO structure is easily crawlable by search engines.

What I Learned Building an Agentic System

The hardest part of building minigames.world wasn't the AI prompting; it was the architecture surrounding it. AI agents don't throw standard errors when they misunderstand a task—they confidently output the wrong format. You have to design your system defensively. Building strict data validation layers and utilizing robust, concurrent backend technologies like Erlang ensures that when the AI acts unpredictably, the system remains stable.

The result? A gaming portal that curates its own portfolio while I focus on improving the engine.

Check out the live results of the agents' work here: minigames.world. I would love to hear how other developers are handling multi-step agent workflows in production!