There is a system to building hit Roblox games. It goes through four phases. We have battle-tested it across games that have collectively reached hundreds of millions of plays. We are now accelerating it with purpose-built AI agents at every stage.
Here is how it works.
Four Phase Development Process
Do people want it?
Player acquisition cost. Opt-in rate for launch notifications.
Is the core play fun?
Session time. Engagement with the core action.
Will players come back?
D1, D7 retention. Shape of the retention curve.
Is it awesome?
Organic growth. Friend invites. Sharing.
Do people want it?
Player acquisition cost. Opt-in rate for launch notifications.
Is the core play fun?
Session time. Engagement with the core action.
Will players come back?
D1, D7 retention. Shape of the retention curve.
Is it awesome?
Organic growth. Friend invites. Sharing.
| Phase | Question | What we measure |
|---|---|---|
| 1 | Do people want it? | Player acquisition cost. Opt-in rate for launch notifications. |
| 2 | Is the core play fun? | Session time. Engagement with the core action. |
| 3 | Will players come back? | D1, D7 retention. Shape of the retention curve. |
| 4 | Is it awesome? | Organic growth. Friend invites. Sharing. |
We benchmark every metric against our own performance data, collected across tycoons, obbies, competitive multiplayer, social experiences, and branded campaigns. The benchmarks tell us what good looks like for each genre, when a concept is worth persisting with, and when it is time to move on.
Phase 1: Do people want it?
Before writing any code, we find out whether players are interested enough in a concept to come back when it launches.
We create game pages for multiple concepts: artwork, descriptions, screenshots. Each page looks exactly like a live Roblox game. We drive traffic through Roblox's ad system. The ads appear in the same format as organic game tiles, so click-through rate is a direct proxy for how the concept would perform in organic discovery.
When a player clicks play, they see the game is coming soon and can opt in for launch notifications. The opt-in rate is the signal. A player who opts in is saying they want to hear when this game is ready. A cheap click tells us the concept attracts attention. A high opt-in rate tells us it holds it.
We run several concepts simultaneously, published under throwaway groups with no brand exposure. For branded games, we are testing whether the concept resonates with Roblox players before the brand's name is attached. The concepts that acquire players cheaply and convert opt-ins are the ones most likely to scale organically.
Game Spark
Generates and fleshes out game concepts using game design fundamentals and data from popular Roblox games. Game Spark encodes what we know about what works on the platform, from genre mechanics to audience preferences, and rapidly produces multiple concepts for testing. Human judgment guides which directions to explore. Game Spark makes it fast to explore many.
Storefront
Creates the game page that sells the concept to players. Storefront generates screenshots informed by visual patterns from successful Roblox games and writes page copy that follows best practices for conversion on the platform. Each concept gets a test-ready game page in hours, so we can run multiple concepts through Phase 1 simultaneously.
Phase 2: Is the core play fun?
We are looking for session time. Players spending minutes with the core action, not seconds. At this stage there are no progression systems, no unlocks, no reason to return tomorrow. We have built only the base layer: the thing the player repeats. If players are spending time with it, the core play is working.
We build a playable prototype. Simple colours, simple shapes, just enough to be playable. No branding, no monetisation. The only thing being tested is whether the core action is enjoyable. Can a player pick it up, understand what to do, and want to keep doing it?
On Roblox, players click and are playing within seconds. No download, no payment, no waiting. A player who has committed nothing has no reason to persist if the game does not immediately make sense. There are thousands of alternatives one click away. The core mechanic has to be understood and enjoyed in seconds. Simplicity at the surface does not mean shallow design. The depth comes later, in the retention layer.
Player acquisition costs pennies. Real players are in a prototype within days generating meaningful data. If the prototype shows engagement, we continue. If it does not, we move to the next concept from Phase 1.
Proto
Builds high-quality playable prototypes in days using long-running agentic coding workflows. Proto encodes our development knowledge and platform expertise, compressing what used to take weeks into rapid iteration cycles. More concepts prototyped and tested, with higher fidelity at each version.
We have tested dozens of concepts through this pipeline. The data from every test, whether it succeeded or not, feeds into our benchmarks. A concept that performs below our threshold for its genre is killed or reworked. A concept that outperforms is fast-tracked.
Phase 3: Will players come back?
We are looking for repeat sessions. Players returning the next day, and the day after, correlated against our retention benchmarks for the genre. A tycoon game has a different retention profile to an obby or a competitive multiplayer. Our benchmarks reflect those differences because they are built from games we have operated across each of those genres.
We add the retention layer: progression systems, unlocks, goals to work towards. The prototype gets polished graphics but remains unbranded. The core loop, the mechanism that drives players through the base layer and the retention layer, is tuned and tested.
This version launches publicly on Roblox. Because it is unbranded, there is no reputational risk. We can launch, gather data, update, and relaunch as many times as needed.
The metrics shift to multi-day retention: D1, D7, and the shape of the retention curve over time. The game stays in this phase until the retention numbers meet our thresholds for the genre. If they do not, the concept is shelved. If they do, the game has earned its investment in branding and polish.
Drop-off Detective
Interrogates player analytics data at a depth that was not previously practical. Drop-off Detective identifies friction points, the specific moments where players disengage, and patterns in how different player segments move through the game. Each round of iteration is informed by targeted analysis rather than guesswork. Built on the same principles as our research agents, applied to game analytics.
Phase 4: Is it awesome?
The unbranded game comes down. Brand identity goes in: custom assets, characters, environments, and narrative. The game relaunches as the official branded experience.
By this point, gameplay is proven fun and retention is proven sticky. The brand enters Roblox backed by data from real players across multiple rounds of testing.
This phase adds the features that drive organic growth: social mechanics, sharing hooks, and the quality of life polish that turns a good game into one players recommend to their friends. These features are not core to the game loop, but they are critical to scaling within Roblox's discovery algorithm.
Roblox's Recommended For You algorithm now rewards engagement quality over raw popularity. In the 18 months since it launched, 13 of the Top 20 most-visited games turned over. New games can find massive audiences if the engagement metrics are strong enough.1
The Minimum Awesome Product is the smallest version of the game where desirability, fun, retention, and brand are all working together, and the experience is good enough that players tell someone else about it.
Sunsilk reached 90M+ visits with no marketing spend. The World Gold Council achieved 89% brand lift. Dettol drove a 35.8% knowledge increase across 16.8 million plays. The process works. The agents make it faster.
