This article uses Shadowverse: Worlds Beyond as a case study for combining AI agents, simulation, robust optimization, and external memory so that one human does not have to remember and test everything manually.
1. “You must be an expert before you can give instructions” is now only half true
The old workflow was simple: research it yourself, implement it yourself, test it yourself.
The emerging AI workflow is different. The human defines the purpose, desired end state, constraints, and acceptance criteria. Then the agents receive the work:
Research it. Produce options. Compare them. Implement. Test. Bring back the results.
The human reviews the output and says:
“Discard A. Use B. This part is wrong. Retest.”
That is already management.
Once the same human also decides which problem deserves time and compute, whether a project should exist at all, and what success means, the role starts resembling executive management or production.
OpenAI’s 2026 description of Codex explicitly emphasizes delegating substantial work to multiple agents, running tasks in parallel, reviewing changes, and redirecting agents. Its AI-native engineering guidance similarly separates work into “Delegate,” “Review,” and “Own”: low-risk execution can be handed off, important outputs are reviewed, and strategy, standards, and accountability remain human-owned.
The structure changes from:
human → task
to:
human → AI employees → task
No employees on payroll.
Still an organization.
2. The CEO’s job becomes less “know every answer” and more “judge whether the answer is good”
Zero expertise is still dangerous.
If an agent announces “Optimization complete!”, someone must still ask why, what alternatives were compared, whether it breaks when assumptions move, whether the test resembles reality, whether the dataset is stale, and what failure costs.
Otherwise:
AI: “Done!”
Human: “Looks good!”
Reality: “It is broken.”
The shift is not from expertise to ignorance. It is from implementation-level knowledge of everything toward enough domain knowledge to evaluate, challenge, and govern outputs.
You do not need every card effect memorized. You do need to notice when an “optimal” deck has a 25% matchup hidden behind a beautiful average.
3. Before playing Shadowverse, build the simulation factory
The normal loop is: start the game, read cards, build a deck, lose, modify it, lose again, play 100 matches, finally begin to understand the metagame.
With computation, the order can be reversed:
- Collect the card database.
- Collect tournament and public decklists.
- Implement game rules and card effects.
- Define mulligan and play policies.
- Simulate tens or hundreds of thousands of games.
- Build a matchup win-rate matrix.
- Extract weak matchups, variance, and brick rates.
- Let the human test only the finalists.
Instead of manually testing 100 decks for 50 games each — 5,000 human games — machines can eliminate most candidates first.
We are writing the strategy guide before installing the game.
There is one small issue: the simulation may take so long that a new expansion arrives before it finishes.
Excellent. We accidentally founded an R&D department.
4. Optimize right before a new pack and your “best deck” may die five days later
As of August 22, 2026, Shadowverse: Worlds Beyond had announced the ninth card set, Revenants of Azvaldt, for release on August 27.
August 22: “Optimal deck complete!”
August 27: “New cards have arrived.”
Optimal deck: “Goodbye.”
The rational pre-expansion investment is therefore not necessarily crafting cards. It is finishing the factory: data ingestion, rules engine, AI matches, aggregation, matchup matrices, robust evaluation, reporting.
Those components survive the expansion.
Wanted to play a card game. Became manufacturing.
5. Robust optimization: not “best on average,” but “doesn’t die when the world moves”
Expected-value optimization chooses the deck with the highest average win rate under the metagame you predict. But forecasts are wrong.
| Deck | vs Dragon | vs Rune | vs Sword | vs Forest |
|---|---|---|---|---|
| A | 70% | 65% | 60% | 25% |
| B | 57% | 56% | 55% | 54% |
If Forest is rare today, Deck A can look amazing. If Forest becomes popular next week:
Deck A has left the building.
Deck B is less spectacular but survives a misspecified metagame.
Robust optimization formalizes this intuition. Instead of pretending uncertain parameters are known exactly, it places them inside an uncertainty set and searches for decisions that remain acceptable under unfavorable realizations.
For a card game, perturb metagame shares, matchup estimates, going-first advantage, mulligan variance, failure to draw a key card, and tactical-policy quality.
A strategy with a peak score of 95 and a floor of 20 may be less attractive than one with a peak of 80 and a floor of 65.
The goal is not the most explosive gorilla. It is the gorilla that still functions when thrown into the wrong zoo.
6. Unfortunately, a perfect simulator cannot press the buttons for you
A machine can find the best deck. The human can still look at a real board and think: “Wait, which card am I supposed to play first?” And lose.
But even this training can be compressed.
Use simulations to extract matchup mulligan rules, high-frequency positions, common lethal patterns, catastrophic misplays, common opponent responses, and decisions with the highest impact on win probability.
Then the human does not train “the entire game.” The human trains the 20% of positions that decide most outcomes.
Deck search: machine. Position mining: machine. Final in-match choice: human.
The CEO still has field training — just not a 100-hour induction course.
7. Is storing every card in human memory actually terrible storage architecture?
Eventually the obvious question appears:
“Why am I paying brain-RAM to store the entire card pool?”
Cognitive science calls the use of notes, devices, reminders, and other external resources instead of internal memory cognitive offloading.
A 2024 computational account modeled the choice in terms of costs: brain-based memory has limited capacity, so keeping an item internally carries opportunity cost; external memory carries a small retrieval/action cost but has effectively much larger capacity.
That maps beautifully to card games.
Do not keep the whole database resident in brain RAM. Use GitHub or a database as storage. Keep only cache-worthy information internally: major archetypes, your win condition, frequently played threat cards, lethal ranges, mulligan rules, and common branches.
The obscure tech card with 3% adoption?
Read it from disk.
8. Worlds Beyond card data can in fact be collected automatically
Public unofficial GitHub projects already demonstrate this.
mariokart761/ShadowverseWB_card_data contains a crawler that retrieves Worlds Beyond card data in five languages and stores card details, skill text, evolved forms, related-card information, set names, and more as JSON.
So somebody is not manually typing hundreds of cards one by one: “Three cost… two attack… three defense…”
The crawler manager is on shift.
Another public repository, ParticleG/shadowverse-wb-db, stores card names, skill text, cost, attack, life, rarity, set information, and images in cards.json.
But one distinction matters enormously:
collecting card text is not the same as executing card behavior correctly in a simulator.
Extracting “Fanfare: deal 4 damage” is comparatively easy. Correctly implementing targeting, timing, buffs, damage reduction, full-board states, and interactions is the painful part.
The card-catalog department is automated.
The rules-engine department is working overtime.
9. Turn the human brain into a decision engine, not a database
Put everything together:
GitHub / DB → long-term memory.
Crawler → data acquisition employee.
Simulator → virtual-experience employee.
Codex / AI agents → research, implementation, testing, reporting.
Human → purpose, evaluation criteria, priority, final decision.
The old model said: “To become strong, memorize all cards and play hundreds of games.”
The new model can say: “Let machines play hundreds of thousands of games and compress the useful conclusions.”
Do not use the executive office as a card warehouse.
Ask the department to bring the file when needed.
10. This is no longer merely a game strategy; it is a one-person-company operating system
Define purpose. Define the desired state. Define constraints. Delegate research. Compare options. Simulate. Choose solutions that survive uncertainty. Execute. Review. Correct. Store knowledge externally.
That works for software, investment testing, process improvement, content production, research, purchasing decisions, and games.
The product is not merely “the strongest deck.” It is a decision factory.
The workers happen to be AI.
The executive does not memorize every operational detail, but must still own what must be achieved, what counts as good, what should not be trusted, and when to say GO.
Wanted to try a card game.
Became CEO before memorizing the cards.
11. Conclusion: design before memorizing; delegate experiments before grinding them yourself
Use the sequence:
research → structured data → simulation → robust evaluation → candidate compression → human playtest
Memory should work the same way:
memorize everything → no
make everything retrievable → yes
cache only frequent, decision-critical information → ideal
The final architecture is simple:
AI plays one million games.
GitHub remembers everything.
The human decides only what matters.
All we wanted was Storm to face.
Somehow the strategy became organizational design.
Before the Storm goes boom,
the management meeting goes boom.
