I built a deckbuilder about working with AI agents, using a fleet of AI agents
by Baran, solo developer · August 2026 · ladder@yazfab.com
LADDER is a roguelike deckbuilder set in the corporate tech world of 2026, where every engineer works beside AI agents, every PM promises AI-native features, and "just use Claude" is both a solution and a curse. You climb the career ladder through card combat. You don't always kill the boss; sometimes you fill a meter, because that is how work actually resolves.
I am a solo developer. I built it with a fleet of AI agents.
That sentence is doing two jobs, and that is the point. The game satirizes AI-era knowledge work, and building it meant living inside the exact dynamics the game satirizes: delegating to agents, reviewing their work, watching them fail in ways that were sometimes hilarious and sometimes expensive, and slowly learning which decisions were safe to hand over and which were not. I did not plan the recursion. It showed up on its own, and by the end it was the most honest design document the game had.
What the game is
Four characters work at the same unnamed company during the same fiscal year: an Engineer, an Engineering Manager, a Product Manager, and the CEO. An enemy in one campaign is the player in another. Each character is a structurally different game. The Engineer levels up. The EM balances AI literacy against politics. The PM races a roadmap clock against stakeholder trust. The CEO starts at the top, and her game is a treadmill: the question is not whether she climbs but whether she stays.
The content adds up to over 600 cards, over 380 enemies, over 230 events, and over 170 relics across the four campaigns.
What broke
If you only publish the wins, nobody should believe you. Two failures, picked for being instructive rather than flattering:
The CEO was unwinnable and nothing noticed. Late in development, the CEO campaign's last two tiers were structurally unwinnable: a resolve check was typed in a way that could never trigger against those bosses' decks. Every automated policy lost, and the harness dutifully reported low winrates, which looked like a tuning problem instead of an impossibility proof. A human reading the sim output with suspicion, not more automation, is what caught it. The fix was small; the lesson was not. Aggregate metrics will happily describe a broken game in the language of a hard one.
The harness called wins failures. Several characters can win without reducing anyone's hit points, by filling a meter, because that is the game's whole thesis about how work resolves. The playtest harness scored those runs as failures for weeks, because its definition of winning was written for the one character who fights. Teaching the tooling that the game had more than one win condition was a bigger job than implementing the win conditions had been.
What stayed human
Direction and taste, all of it. Which of five drafted card batches ships. Whether a joke lands or just references. What the CEO's game is actually about. Code review and every merge. Store accounts, submissions, and the legal boilerplate. And the decision, over and over, of what not to build.
The division of labor that emerged was not "AI does the easy parts." It was: agents do everything that has a checkable definition of done, and I own everything that does not. The game itself ended up being about the same boundary, which is either a coincidence or the least surprising thing in the world, and I no longer think it is a coincidence.
Where it lands
LADDER is finished and heading for the stores, on iOS and Android. It is not published yet, so there is nothing to download today and no store link to give you. When there is one, this page and laddergame.ai will carry it.
If you have questions about the setup, the harnesses, or what I would do differently, ask. That conversation is half the reason this post exists.
LADDER is not affiliated with or endorsed by Anthropic, OpenAI, or any AI vendor named or parodied within it.