The most revealing user of Ryan Endacott's software may be the one that confidently invents a command. While testing the developer kit for Star, his AI game-building platform, Endacott gave more than 30 fresh agents the same assignment: build a game. They fabricated identifiers, reached for functions that did not exist, repeated the same leaderboard bug, and once called an API from a game loop roughly 60 times every second. It was less a usability test than a small carnival of synthetic certainty.
Endacott kept watching. A placeholder that humans understood as an instruction looked to an agent like an invitation to make up a value. A nonexistent command was guessed so often that it began to look like the command the product should have offered all along. The looping API call suggested that safety should live inside the tool, not inside a hope that every caller will behave sensibly. Each failure became a clue.
This is the current chapter of a career spent lowering the drawbridge around complicated systems. At the University of Missouri, Endacott helped turn a quiet computing club into a busier community and helped establish the campus hackathon that became TigerHacks. At Loon, he worked on Mission Control software for a fleet of high-altitude balloons carrying internet connectivity. SolStar tried to make crypto less forbidding inside Discord. Stella let people converse with websites instead of excavating their menus. Star now lets an idea arrive as a sentence and leave as a playable browser game.
Before the balloons, a room full of students
Endacott studied computer science and mathematics at Mizzou. Those years included a 4.0 GPA, scholarships, teaching work, open-source projects, and internships at Garmin, Google, Facebook, and Clever. The more useful detail is social. As a freshman, he ran for president of the campus Association for Computing Machinery when the organization was relatively inactive. He and the other officers organized workshops, encouraged side projects, and increased typical meeting attendance from roughly 20 people to 50.
The hackathon followed. A weekend event can look like pizza arranged around laptops, but its actual product is permission. Students who have spent a semester submitting assignments suddenly build something unnecessary, funny, ambitious, or all three. Endacott has described watching students shift from course-takers into creators. That conversion still seems to animate his work.
His own student projects had the same playful utility. CrowdArmor, a real-time social-media security project, won a prize at Hack the Midwest in 2013. His Ruby version of Verbal Expressions turned the punctuation thicket of regular expressions into readable instructions and attracted hundreds of stars on GitHub. He also built a toy Scheme interpreter because sometimes the direct route to understanding a language is to make a small one yourself. None of these projects needed a corporate strategy. They were proof that curiosity works better when it can compile.
Mission Control, with humans included
Loon was a marvelous engineering noun: a fleet of stratospheric balloons designed to bring internet service to remote places. Endacott joined in 2016 and worked across its Mission Control stack. He built interfaces for flight engineers, tools for inspecting telemetry, systems for validating hardware before launch, and map overlays for weather and steering conditions. These were not speculative demos. A flight team needed to read the software, trust it, and act.
He later wrote warmly about dawn launches and the colleagues who taught him what kind people could do together on difficult problems. When Loon shut down in 2021, the fleet disappeared but the lesson traveled. Complex systems become useful when they meet people at the point of decision. The interface is not garnish on the machine. It is where the machine joins the world.
“To inspire you and show that anyone can build with AI.”Ryan Endacott, writing in 2025
Three names, one preoccupation
After Loon came SolStar, a tool for sending crypto inside Discord communities. Endacott and co-founder Drew described it as a kind of Venmo for Discord, an intentionally familiar comparison for an unfamiliar financial system. The startup entered Y Combinator's Summer 2022 batch. Crypto's winter arrived, but Endacott's interest in turning intimidating machinery into something conversational survived.
Stella pointed that instinct at the web. Instead of expecting visitors to parse a site's architecture, its AI concierge let them ask for what they needed. A Town of Carmel deployment answered resident questions about ordinary civic tasks, the sort of information that can technically be online and still feel hidden. Endacott later led an AI workshop for more than 60 New York town officials and wrote a practical guide for municipal leaders. His pitch emphasized time saved and information made easier to reach, not a robot wearing a mayoral sash.
Star is the latest transformation of the same Y Combinator company. Its public promise is plain: describe a game, iterate, and play it without first becoming a programmer. The platform sits in a lively new category of AI creation tools, but Endacott's version carries the memory of those Mizzou rooms. The important moment is when a person with an idea realizes the technical door is open.
When the user is an agent
Star's SDK pushed Endacott into a newer design problem. Human developers know that YOUR_GAME_ID is a placeholder. An AI agent may cheerfully replace it with plausible nonsense and continue. Humans tend to avoid expensive calls inside rendering loops. An agent has no bodily intuition for a server bill. Documentation written as a sequence may be scanned as a bag of patterns.
Endacott responded by changing the product's grammar. Placeholders point toward the file containing a real value. Repeated function calls return safely. Guessed commands become aliases when the guesses reveal a natural convention. Documentation can be installed directly into an agent's environment. The philosophy is generous and unsentimental: observe the user you have.
When several independent users make the same mistake, stop treating it as a coincidence. Repeated confusion can reveal a missing command, a leaky abstraction, or an unsafe default. Fix the interface at the point where the misunderstanding keeps gathering.
There is a pleasing circularity here. The college organizer who created a better setting for students to learn is now creating a better setting for agents to operate. The populations differ rather dramatically. The method does not. Watch behavior. Remove avoidable friction. Turn common failure into shared infrastructure.
A beautiful result meets an audit
Endacott's curiosity also wanders beyond products. In 2025 he asked an AI system for a scientific mystery that might be explored with public data. It proposed recurring pilot-whale strandings. The first analysis appeared to confirm a magnetic-field hypothesis perfectly. It was also built on fabricated and badly handled data.
He did something rarer than obtaining a perfect result: he published the failure. With a newer model and verifiable public data, the project audited its numbers, found incorrect coordinates and magnetic measurements, and tested the original idea again. The tidy hypothesis died. The work then moved toward environmental signals and an early-warning experiment, with its limitations left visible.
The episode supplies a counterweight to Endacott's enthusiasm for AI-assisted building. He has written that AI produces his code and lets him work much faster. Yet the science experiment shows the tax on that speed: provenance matters, tools break, and an elegant answer can be a trap. Optimism becomes credible when it comes with an audit trail.
The useful founder habit is not always knowing. It is making uncertainty cheap enough to test.A pattern across Endacott's public experiments
The invitation keeps getting wider
Endacott calls himself a builder and tinkerer. He angel-invests in founders, maintains open-source code, and has accumulated a charming cabinet of side projects: a toy Scheme interpreter, a tiled-background viewer, a multiplayer fish simulation, and an AI science repository. The variety makes more sense when viewed as a single temperament. He likes finding out what a system will let him do, then moving the boundary.
Star's game maker is the commercial expression of that temperament. Its audience is not limited to professional studios. It includes the person who can explain a game mechanic, a mood, or a ridiculous premise and would like to see it move on screen. In that exchange, language becomes a rough new programming layer. The code still exists. It simply stops being the price of admission.
There are grand predictions available about where AI creation goes next. Endacott's work suggests a more practical measure: count the people who move from watching to making. That was the promise of a campus hackathon. It was the human requirement inside Mission Control. It shaped Stella's conversational websites and now Star's prompt-built games. The technologies keep acquiring new names. The invitation remains recognizably his.