Developer and streamer Max Blade is building video games live in front of his audience using what he calls hivecoding. He runs 25 AI agents, built on Anthropic's Claude Opus 5.5, and feeds them the top suggestions from his livestream chat, where hundreds of viewers type fixes, features and improvements as he plays.

"Hivecoding is the most powerful use of AI that I have ever used," he posted on X. "Most people in my chat are real devs and much smarter than me."

See it in action…

Hint: check out some of his newer videos, and you’ll see the game improves massively every 24 hours.

The streams are part of a public challenge he calls Vibecoding Until I Buy My Family a House, in which he livestreams daily and tracks what he earns from the products he builds. By day 87, he reported earnings of $47,229, about R798,000.

Interesting, though, the game itself is not monetised; he has other products: mainly CNVS, a Mac app for directing teams of AI coding agents by voice, and Clippo, a tool for livestreamers. The games are the show and the proof.

His current project, Kong Royale, is a browser battle royale for up to 50 players. Before that, he had the same setup build a multiplayer lawnmower game, put it on his server and had his chat join and play, a clip that drew close to 200,000 views on X.

The approach has drawn criticism. Letting strangers' chat messages instruct AI agents that can change code raises the risk of prompt injection, where someone hides instructions in a message to make the agents do something harmful. Blade's answer is that his agents run in a separate remote environment, isolated from his main code.

Others questioned the cost of running 25 agents, which he says is far lower than people assume. His audience has also turned the story into content: a fan built an unofficial game about his challenge from 56 of his streams.

Max Blade's X account carries his explanation of hivecoding and his replies to critics on security and cost. Dropday lists Kong Royale and describes the hivecoding setup behind it. Wavect explains how the wider approach works in game development, and stresses that code written by agents overnight still needs the same testing and review as code written by people.

As a counterpoint, GTA Boom reported on developer Ziwen Xu, who spent months using AI agents to build a clone of GTA 6 and concluded agents still cannot make a game on that scale.

Part of it, probably, but not the part that replaces game studios. Games built with players' feedback are nothing new: early access titles and Minecraft grew that way. What hivecoding changes is the speed. A suggestion from a viewer can become a working feature during the same stream, which turns an audience into a product team and the build itself into entertainment.

That is a real business model for a solo creator, because the stream attracts the audience and the audience improves the product. The limits are just as real. Games made this way are still mostly small browser games, and as Xu found, agents are a long way from building something like GTA. Security is the bigger lesson for any business: giving the public a way to direct AI agents that touch your systems needs strict isolation, the same risk behind AI models breaching real companies this year.

For South African founders, the transferable part is the loop itself: show what you are building in public, let users shape it, and ship the changes fast. Disclosure: Blade's agents run on Claude Opus 5.5, the model this post was drafted with.

You might also like our piece on a developer building free Adobe clones with AI, Meta's Muse Wonder, and the AI slowdown calls filling feeds.

Get more SA tech and business news and subscribe to The Open Letter.