AGI Inc. — Build AI, Feel the Consequences

AGI Inc. — build AI, feel the consequences

AGI Inc. is a playable AI-company simulator that teaches the decisions behind taking an AI product from first compute to a customer-facing launch.

Most AI learning tools stop at a notebook, a benchmark, or a dashboard. This game makes the trade-offs tangible: the player walks an office, buys and assembles compute, decides whether to rent capacity for cash or bank it for training, audits training evidence, trains a support-ticket model, red-teams it, makes a release decision, and then sees the customer and operational consequences.

The problem

AI systems are often taught as isolated model choices. In practice, data quality, compute allocation, evaluation, release discipline, customer trust, latency, revenue, and incident response are one connected system. AGI Inc. turns that system into a short, understandable learning loop.

What the player does

Build the infrastructure — order components, assemble a working rack in a spatial RIG DROP puzzle, and deploy it. Make the business trade-off — move capacity between immediate rental income and banked training compute in the Operations computer. Audit the evidence — repair four realistic data-quality failures: privacy leakage, duplicates, missing coverage, and evaluation leakage. Train deliberately — configure data, model size, and learning rate in Model Lab; watch live training and validation curves; stop before the candidate overfits. Red-team before release — classify four observed failures, including a safety-critical case, then make a ship-or-hold decision. Own the outcome — inspect the Support Inbox and operate the resulting live product through customer impact and incidents.

The same deterministic game state carries every choice through the whole simulation. A data choice changes training; the training result changes evaluation; the release decision changes the customer and operations outcome.

Who it is for

AGI Inc. is for junior and mid-level AI builders, product owners, and risk or safety practitioners who need to feel the operational trade-offs behind an AI release, not only read about them in a dashboard.

What a player practices

A single session lets the player trace how weak data or an unstable training choice can surface as a support failure, weigh near-term compute revenue against training capacity, and treat red-teaming and release discipline as product decisions rather than checkboxes.

Why now

Powerful models make it easier than ever to ship an AI feature quickly. AGI Inc. is a safe simulator for understanding the system-level consequences before those shortcuts reach customers.

A 3-minute judge path

Run the app and open /?demo=1. It starts with a real deployed rack, audited data, and banked compute so the judge can inspect the complete learning loop without replaying setup.

Open the ready Model Lab block inside the Floor 1 Operations computer. Train a candidate and watch the live loss curves. Red-team the candidate, make a release call, and inspect the customer-facing result.

This is a shortcut into the real game state, not a mocked video flow. The full route at / includes the walk, shopping, delivery, rack build, data audit, compute allocation, and the same Model Lab → evaluation → launch loop.

How Codex and GPT-5.6 are used

Codex was the development environment for the project: the Three.js office, TypeScript game systems, campaign runtime, save/pause UX, deterministic browser tests, judge path, recording harness, and documentation were iterated in Codex sessions.

GPT-5.6 powers the optional Live ARIA co-founder through the OpenAI Responses API, using gpt-5.6-luna. With a player-supplied API key, ARIA can explain the current decision surface, trade-offs, nearby interaction, or route. It receives only a read-only game snapshot, keeps that working context private, and cannot click controls, grant resources, or alter game state. The game remains fully playable without an API key.

That boundary is deliberate: the player makes the consequential decision; GPT-5.6 helps them understand it. For example, ARIA can explain why a chosen learning rate makes the curves unstable and suggest what to inspect next, but it can never press Train or Ship for the player.

Core implementation was iterative across several Codex sessions. The /feedback Session ID in this submission points to the primary implementation session, which contains the Campaign Director, Floor 1 learning loop, persistent state, physical navigation, and browser-game validation. Focused follow-up sessions covered RIG DROP and final polish.

Built for honest learning

The support-intent taxonomy references BANKING77. The ticket examples, loss curves, scores, and routes are original deterministic teaching simulations, not claimed benchmark results. The project includes unit tests and browser-game tests for the playable route.

Run it locally npm install npm run dev

Open http://localhost:5173/ for the complete game or http://localhost:5173/?demo=1 for the short judge path.

Repository: https://github.com/gagarinyury/agi-inc