When Junyoung Kim arrives at Bagelcode, he wants the blank page to have disappeared. The mobile-game company's AI agents have spent the night inspecting issues, drafting code changes, and lining up proposals. A developer can begin the morning with a question sharper than “What should I do?” The new question is “Is this any good?” It is an appealing rearrangement of office life: the machine gets the graveyard shift; the human keeps judgment.
Kim, Bagelcode's co-founder and co-CEO, has been preparing for this reversal for longer than the current AI boom makes obvious. His company was built on the conviction that behavior leaves clues. Players click, return, buy, ignore, and churn. A studio that records those decisions carefully can learn not only what worked, but why the next attempt deserves a budget. Bagelcode spent more than a decade joining game design, live operations, creative production, marketing, and data into a single feedback loop. AI did not invent that loop. It simply arrived hungry enough to eat the whole thing.
A useful miss
Bagelcode began in 2012 with five developers, including Kim and fellow co-founder Ilhwan Yoon. Their early path carried none of the tidy inevitability that founder profiles acquire in hindsight. The team tried a real-time puzzle game for Kakao. It did not become the business. Yet one small mechanic inside the game - a random slot machine - attracted spending. The failed product had left behind a behavioral note in the margin.
The founders followed it toward social casino, a category aimed largely at Western players and unfamiliar to a small Korean team. Their first publishing arrangement fell away. They reserved part of their early investment for travel, assembled lists of major US publishers, cold-called them, and hunted for encounters at conferences. The pitch was bracingly direct. When one prospective partner suggested buying the company, Kim's answer was effectively: not cheaply, not now; teach us to grow and perhaps we can discuss a more expensive future.
Big Fish Games became that partner. Kim later recalled the day an $8 million payment landed in Bagelcode's account. The number mattered, but so did the education attached to it. Bagelcode learned how a US publisher evaluated a game, how a foreign audience behaved, and how distribution could shape the product rather than merely carry it. The company was not born global by putting pins in a map. It became global by accepting that the missing expertise might live elsewhere.
“Creating revenue from zero to ₩10 billion is art. From ₩10 billion to ₩100 billion is science.”Junyoung Kim, translated from a 2022 interview
The art, the science, and the invoice
Kim's art-and-science distinction is the shortest route into his operating philosophy. The first stage demands a leap: a world, a mechanic, a feeling, something that did not exist yesterday. Data has little to say about an unmade thing. Once a game finds genuine demand, however, growth becomes a discipline of repeatable questions. Which players return? Which offer lands? Which advertisement earns back its cost? Which market has room for another experiment?
the bet
to players
behavior
the lesson
Bagelcode paid tuition in media spend. Kim once joked to Yoon that after spending roughly ₩10 billion on marketing, he thought he understood it. By the time the total had reached about ₩100 billion, the view had changed again. New advertising channels required actual experiments, not confident conference-room opinions. A test large enough to produce a meaningful signal could itself cost hundreds of millions of won. Knowledge arrived with invoices.
The company also bought learning at corporate scale. In 2018 it acquired the social-casino operation of the UK's JPJ Group. The appeal extended beyond titles such as Jackpotjoy and Bingo Lane. The acquisition brought a portfolio, experienced people, and a deep user database. During the nine-month process, Kim said, Bagelcode's own revenue grew almost tenfold, shrinking what had initially been a startling size gap between buyer and target. Funding supplied the purchase price; operating momentum changed the psychology.
Offices and teams followed in places where relevant talent already lived. Post-merger work led to a UK presence. Tel Aviv offered veterans of the social-casino and advertising industries. Teams also formed in Dnipro and Hanoi. Kim and Yoon discovered that language and time zones were not the only difficulties. Work cultures carried different assumptions about disagreement, authority, and what a polite yes actually meant. Their company had to learn the people as deliberately as it learned the players.
Let reality cast the vote
The system eventually produced less romantic numbers. Bagelcode reported ₩81 billion in 2021 revenue and turned profitable in the first half of 2022. It received Korea's $50 Million Export Tower and a Service Tower later that year. For 2024, the company reported ₩111.7 billion in revenue and ₩13.8 billion in operating profit. Club Vegas and Cash Billionaire remained important, while the team pushed into casual, defense, arcade, strategy, and idle-game experiments. A reliable hit is a lovely asset. It is also a concentration risk wearing party clothes.
The transferable part of Bagelcode's story is not the slot machine. It is selection. Creative businesses generate more possibilities than they can responsibly fund. The advantage goes to the organization that can place many modest bets, observe honest behavior, and abandon a beloved weak idea without holding a funeral for it.
That philosophy made Bagelcode's 2026 game jam especially revealing. Seven days produced 124 finished mobile games, compared with nine at an event three years earlier. Preparation shrank from 60 days to seven. Individual entrants, once absent, represented 88 percent of participants. AI tools and the company's GameBakery.ai engine had made production startlingly abundant.
Abundance creates its own nuisance: someone must choose. Bagelcode's answer was to release the games and let player retention help determine the winner. Executives could admire a pitch. Designers could defend a flourish. The returning player got the final vote. It was the original failed-game lesson, now conducted at industrial speed.
The colleagues who do not need coffee
Kim's response to generative AI was characteristically operational. Bagelcode paid for employees' ChatGPT subscriptions early, then expanded support to more than 50 AI tools. His reasoning was domestic before it was corporate: someone who never uses AI in ordinary life will not become fluent simply by arriving at work. Regular internal events let developers, artists, and other specialists show one another what worked. Kim concentrated first on executives, demonstrating results until they could carry the practice into their own departments.
By 2026, the ambition had moved from AI-assisted to “AI-agent-driven.” In Kim's definition, an agent combines data, code, and a language model, then acts toward a goal instead of waiting politely in an empty chat window. Bagelcode prepared dedicated Mac minis as isolated sandboxes, giving agents room to work while reducing the risk that experimental automation could reach sensitive systems. The gesture is faintly comic - a second computer for a colleague without a chair - and entirely consistent with Kim's habit of making abstract strategy physical.
“An agent does the work and first asks, ‘Is this right?’”Junyoung Kim, translated from a 2026 interview
The new productivity promptly broke an old ritual. A development operation accustomed to around ten daily code commits began seeing 100 to 200. Recording and sharing the changes threatened to consume more time than making them. Bagelcode replaced the overwhelmed process with an internal app called ToDos, which can report completed work and propose what comes next. Artists and planners began tuning it to their own teams. The company did not merely automate Jira. It discovered that Jira's assumptions belonged to a different rate of production.
Team size came under inspection too. Kim described a three-person group that requested a fourth member, then withdrew the request. Another human would add documentation, meetings, and persuasion. More agent capacity met the immediate need with less coordination. It would be foolish to turn one anecdote into a universal theory of employment. It is more useful as a reminder: when individual range expands, the old relationship between headcount and throughput stops being arithmetic.
Taste gets promoted
Kim does not describe games as software with decorations. He calls the work a total art connecting planning, visuals, development, data, operations, and customer response. AI can widen each specialist's reach, but the durable advantage, in his telling, comes from deep domain knowledge meeting that amplification. The machine makes the field larger. Expertise keeps the ball in play.
This is where the Bagelcode story becomes more interesting than a familiar tale about doing more with less. Cheap production does not remove the difficult decisions. It floods them. A studio that can create 124 games in a week now needs a sharper test, a clearer portfolio, and adults willing to cancel 123 plausible dreams. When the first draft costs almost nothing, selection becomes expensive in attention.
Kim's career has been an education in that expense. The early team read a clue inside a miss. The founders chose foreign expertise over local comfort. They spent enough on marketing to be corrected repeatedly. They bought a larger operation for its accumulated knowledge. Now they give machines the first move and reserve the human move for review, context, and responsibility.
The aspiration is broader than a faster slot studio. Bagelcode wants to diversify its genres and build a more automated, intelligent method of game production for a global audience. Whether those new games find players remains, properly, unresolved. Kim has arranged his company so the answer will not come from a slogan. It will arrive as behavior: a tap, a return, a quiet decision to play again tomorrow. Even in an office full of agents, the human on the other side of the screen still has the last word.