AGI Tycoon is a management sim about the economics of an AI lab: compute, pricing, funding rounds, and rivals who make their own decisions.
It is also where most of my rules for working with agents came from, because it is the project where I let them do the most.
How it was built
The engine is pure and seeded, with one invariant every change must keep: the value before, plus the sum of the contributions, equals the value after. That makes a wrong number easy to trace.
A headless harness plays the game on its own across a matrix of difficulties. Balance changes are gated on those recorded runs. A telemetry catalog describes every field the game records, and runs are written to SQLite so the agent and I look at the same evidence.
The repo’s rules for agents are short: write the test first and confirm it fails, use one implementer and at most one reviewer, and never push or deploy without my approval.
The build you can play here is from July. Version 0.9 is further along and not yet published.
- Status
- Live
- Commits
- 392
- Test files
- 495
- Codex threads
- 634
- Built with
- React, TypeScript, Vite, Zustand, SQLite telemetry