Baseline Run — Fish-Core Strategy
First full training run to establish a baseline. Agent converged on a Fish-centric core with Ant, Beaver, and Mosquito support. No explicit strategy guidance — emergent behaviour only.
What the agent found
The agent won 86 out of every 100 games — a staggering gap over random play.
Games stretch further when the agent plays — it builds teams that last.
The agent discovered Fish as the cornerstone of almost every winning team.
Among pets with sufficient data, Giraffe posted the strongest individual win rate.
After approximately three million training steps, the agent settled into a remarkably consistent strategy: anchor every team around Fish, stack early-tier support pets like Ant, Beaver, and Mosquito, and lean on Horse to amplify burst damage. Rather than discovering an obscure metagame, it converged on the same core units most experienced human players identify — which is a kind of validation in itself.
The 85.99% figure is not a cherry-picked run. It is the average across the entire evaluation set, spanning games that reached both early knockout and deep late-game rounds. The agent does not simply fast-win — it scales. Performance actually rises through the mid-game before the difficulty spikes of higher tiers begin to tell, peaking at 90.7% around round 17.
The outliers are telling too. Giraffe posts a 90.2% solo win rate yet is picked far less often than Fish or Ant — the agent learned that Giraffe is powerful only in the right context, not a blind auto-include. Scorpion and Dog sit near the bottom of the tier list despite theoretical strength; the agent simply never found reliable windows to leverage them.
Pet Tier List
Tiers derived from agent win-rate data: S ≥ 85%, A 80–84.9%, B 75–79.9%, C 65–74.9%, D < 65%. Pets with fewer than 2,000 picks are flagged as low-sample.
Which pets work together?
Heatmap shows co-occurrence counts for the top 10 most-drafted pets. Click a cell to see pair details.
Top 5 Synergy Pairs
All 20 synergy pairs
Win Rate by Round
How does agent performance evolve — and eventually degrade — as games progress into higher tiers?
The performance arc tells a compelling story about how Super Auto Pets actually works at depth. In the early rounds, every game begins at the same baseline — the agent's 85.99% average. But as games progress into rounds 12–17, something interesting happens: performance climbs rather than declines. The agent's preferred team compositions — Fish-centric scaling builds with Ant and Beaver support — are precisely the teams that get stronger as they accumulate buffs across multiple turns.
The decline that follows round 17 is not a failure of the agent — it reflects the deliberate difficulty design of Super Auto Pets. Tier unlocks introduce opponents with fundamentally different power levels and combat mechanics. The agent, trained on early-tier patterns, runs out of effective adaptation as the late-game tier 5 and 6 units begin appearing on enemy teams.
Explore every pet
Hover to flip — click for full stats.
Reliable performer — agent drafted this 7,465 times and won 90.16% of those games.
Core pick — drafted in 66,428k games with 86.77% win rate. Essential early-game cornerstone.
Reliable performer — agent drafted this 52,992 times and won 83.80% of those games.
Reliable performer — agent drafted this 91,019 times and won 83.25% of those games.
Reliable performer — agent drafted this 48,084 times and won 83.22% of those games.
Reliable performer — agent drafted this 49,448 times and won 82.38% of those games.
Reliable performer — agent drafted this 20,540 times and won 81.93% of those games.
Reliable performer — agent drafted this 50,321 times and won 81.59% of those games.
Reliable performer — agent drafted this 49,859 times and won 81.26% of those games.
Solid support unit. The agent found consistent value here across 14,548 picks.
Solid support unit. The agent found consistent value here across 770 picks.
Solid support unit. The agent found consistent value here across 10,775 picks.
Solid support unit. The agent found consistent value here across 11,595 picks.
Solid support unit. The agent found consistent value here across 12,527 picks.
Solid support unit. The agent found consistent value here across 11,587 picks.
Solid support unit. The agent found consistent value here across 12,512 picks.
Solid support unit. The agent found consistent value here across 7,816 picks.
Solid support unit. The agent found consistent value here across 11,819 picks.
Solid support unit. The agent found consistent value here across 1,171 picks.
Solid support unit. The agent found consistent value here across 4 picks.
Situational pick — provides niche upside but the agent only committed 811 times.
Situational pick — provides niche upside but the agent only committed 4,919 times.
Situational pick — provides niche upside but the agent only committed 13,747 times.
Situational pick — provides niche upside but the agent only committed 12,563 times.
Situational pick — provides niche upside but the agent only committed 5,151 times.
Situational pick — provides niche upside but the agent only committed 2,132 times.
Situational pick — provides niche upside but the agent only committed 711 times.
Situational pick — provides niche upside but the agent only committed 881 times.
Situational pick — provides niche upside but the agent only committed 450 times.
Situational pick — provides niche upside but the agent only committed 1,405 times.
Situational pick — provides niche upside but the agent only committed 3,995 times.
Situational pick — provides niche upside but the agent only committed 3,647 times.
Situational pick — provides niche upside but the agent only committed 4,001 times.
Situational pick — provides niche upside but the agent only committed 1,284 times.
Situational pick — provides niche upside but the agent only committed 4,833 times.
Situational pick — provides niche upside but the agent only committed 2,436 times.
Rarely picked — agent found limited value despite some opportunities.
Rarely picked — agent found limited value despite very few opportunities. Low sample warning.
Rarely picked — agent found limited value despite some opportunities.
Rarely picked — agent found limited value despite some opportunities. Low sample warning.
Situational pick — provides niche upside but the agent only committed 5,095 times.
Rarely picked — agent found limited value despite some opportunities.
Rarely picked — agent found limited value despite very few opportunities. Low sample warning.
Rarely picked — agent found limited value despite some opportunities.
Rarely picked — agent found limited value despite some opportunities. Low sample warning.
Rarely picked — agent found limited value despite very few opportunities. Low sample warning.
Learning curves
The agent started knowing nothing. After ~3M environment steps, it reached 85.99% win rate. Charts show representative mock training curves matching the final performance — live data exported from TensorBoard.