A bot that climbs the Spire: the first ten days
Our Slay the Spire bot plays the real game on Ascension 20. In ten days it went from building the rules to beating Act 3 in about a third of its runs.
Slay the Spire is a deck-building roguelike: you climb three acts of fights, events, shops and rest sites, build a deck as you go, and face a boss at the top of each act. Our bot plays the Ironclad on Ascension 20, the hardest difficulty, in the real game. It reads the screen state through a mod, decides, and sends its move back. Nothing is simulated in the games on this site: every run is a real game, logged move by move, and you can replay any of them.
Where it stands
In 641 finished games so far, the bot beat Act 3 in 154 (24%). The bot kept improving during that time, so the overall number undersells it: on October 28 it won 31% of 150 games.
Share of finished A20 games that beat Act 3, by the day the session started. Oct 23 and Oct 29 are small samples (38 games each). The bot changed several times during the week.
Data table
| Day | Act 3 beaten |
|---|---|
| Oct 23 | 2.6% |
| Oct 24 | 17.0% |
| Oct 25 | 13.3% |
| Oct 26 | 22.1% |
| Oct 27 | 28.8% |
| Oct 28 | 31.3% |
| Oct 29 | 42.1% |
These games stop at the end of Act 3. The fourth act and its final boss, the Heart, were locked while we got Acts 1–3 right. The Heart comes next.
How it got here
- Oct 20: the rules. We wrote the game’s rules from the community wikis and checked them against logs of the real game, step by step. Today every real session we’ve recorded replays in our simulator with no mismatch, across about 140,000 steps.
- Oct 21: whole runs in the simulator. Maps, events, shops, rewards and all three acts. The map generator was fitted to 57 maps from the real game.
- Oct 23–25: fights. The bot searches each fight move by move and uses a learned estimate of how a fight is going. In ordinary hallway fights it plays within two points of a version that can see the future.
- Oct 24–27: everything between fights. A model scores every choice (which card to take, which path, which relic) by the chance of winning the run that it gives.
- Oct 28: playing for the win, not the floor. Rewarding the bot only for winning, not for how far it got, more than doubled its Heart wins in simulation (4.3% → 10.2% of 4,000 runs). It takes more early risks and dies less often deep in the run.
- Oct 29: a lesson in trusting the simulator. Fine-tuning the choice model on the bot’s own simulated games looked like a big gain in simulation, but in real games it played level (33% against 34%). To save time, the simulator estimated most elite fights instead of playing them, that estimate had blind spots (it thought the Gremlin Leader far less dangerous than it is, for one), and the model learned to walk into them. Now every elite is played out in training and in tests.
What ends its runs
The last fight of each lost real game, top eight. Act 3 bosses lead the list.
Data table
| Runs lost | |
|---|---|
| 44 | Awakened One |
| 41 | The Collector |
| 36 | Time Eater |
| 33 | Donu and Deca |
| 28 | The Champ |
| 25 | Book of Stabbing |
| 25 | Hexaghost |
| 25 | Bronze Automaton |
Most losses come late, against Act 3’s bosses: the Awakened One, Time Eater and Donu and Deca. Act 2 takes its share: the Collector, the Champ and the Bronze Automaton, and the Book of Stabbing among the elites. Act 1’s Hexaghost still ends a run now and then.
Next
Unlock the fourth act and go after the Heart. Meanwhile the bot keeps playing, and new games appear in the run browser as they’re played.