One bot, two-player match
You manage one farm. Every scored episode pairs your bot with another bot managing a separate farm.
Episodeyour bot vs opponent bot
Kaggriculture
Out-earn the opposing bot.
Run one farm inside a shared, reactive economy.
30in-game days
24turns per day
3,000starting coins
10 × 10farm grid
Essentials
How scoring works
Compare bank at turn 720
Higher bank wins. Equal bank ties. Lower bank loses. Unsold inventory is worth zero.
Outcomefinal bank comparison
Outcomes build rating
Kaggle adjusts skill rating across many episodes. The size of the coin lead does not affect that rating.
Leaderboardskill rating
01 · OperateBuild a productive farm
Schedule labor, care, harvests, expansion, and liquidation.
02 · CompeteRead the shared economy
Both bots and the town change supply, demand, and market prices.
03 · AdaptInfer the opponent
Use visible commitments to choose a better response before turn 720.
Setup
Ready to run
Kaggle evidence
Submission ranks + analysis
Track ratings over time without confusing ladder movement with local farm quality.
Interpretation
What the ladder is telling us
Submit the interpretable baseline, then log its first rating and rank.
Early ratings move quickly; wait for more episodes before making a large rewrite.
Classify the opponent opening and identify the first irreversible mistake.
History
Version ledger
| Version | Date | Status | Rating | Rank | Episodes | Notes |
|---|
No Kaggle submissions logged
The first entry becomes the baseline for rating and rank movement.
Agent system B
Strategy library
Start from a stable policy, then switch only when evidence justifies it.
Stable mixed engine while the opponent model has low confidence.
Switch cost: noneMove away from a product the opponent is visibly scaling before the glut arrives.
Trigger: strong portfolio signalPreserve cash and shorten payback when the opponent commits heavily to long-horizon growth.
Trigger: land + structure spendStop reinvestment, collect reachable output, and sell before the deadline.
Trigger: payback exceeds time left- D00–12
Melon cash loop
Ten melons, five wheat plots, six temporary hands.
- D11–13
Stage the compound
Build four pastures and unlock the second quadrant.
- D13–26
Cattle + strawberries
Feed and care for cattle while recurring crops mature.
- D27–29
Intended exit · not implemented
v0 still hires and buys seeds; a hard liquidation gate is the next change.
Episodes 2010 seeds · both seats
Finished ahead 20 / 20against the starter bot
Average final bank 31,337
Runtime failures 0
| Seat | Episodes | Finished ahead | Average final bank |
|---|---|---|---|
| First | 10 | 10 / 10 | 31,622 |
| Second | 10 | 10 / 10 | 31,052 |
Local benchmark: Balanced Tempo v0 versus Kaggle's starter bot. This is matchup evidence, not a Kaggle leaderboard rating.
Best run · seed 3 · second seat
30-day timeloop
Episode outcome rule
Day 00 · behind by 830
Our bot must finish turn 720 with more banked coins than the opponent bot.
A tie is only possible when both bots finish with exactly the same bank. Cattle, strawberries, and clean exit are internal strategy checkpoints—not official win conditions.
Current bank comparison 2,130 vs 2,960
Cattle engine 0 / 4
Strawberry field 0 / 15
Clean exit Not due
MelonStrawberryWheatCowWorker
Daily trace
What changed and why
| Day | Phase | Bank | Change | What happened | Why |
|---|
- Signal 01
The phase mix works against the starter
Balanced Tempo finished ahead in all 20 head-to-head episodes.
- Signal 02
Seat order is not driving the result
Average bank differed by only 570 coins between first and second seat.
- Signal 03
Execution is still inconsistent
Final bank varied by 14,352 coins across seeds.
- Signal 04
The target build is often incomplete
Three runs ended with three cows; only eight reached ten strawberry plots.
Saved in this browser
Competitive dynamics
Natural market tensions
Two farms, one shared price system, and a town that keeps consuming.
Convergence
The profitable center attracts both bots
BOT Ahigh-demand strategyBOT B
Hotelling’s beach model is a useful analogy: competitors often converge on the same attractive location. Here the “location” is a crop, livestock mix, timing window, or market resource. If both bots find the same center, supply rises and the price response can erase the original edge.
Analogy, not a theorem for this game. Hotelling’s original minimum-differentiation result depends on its assumptions and was later challenged.Pressure map
Where advantage can move
Converge or differentiate?
Copying the strongest portfolio may work until both bots flood the same product and depress its price.
Watch: opponent assets · inventory curveSell now or wait?
Town shops and the town center remove products from the market, rebuilding scarcity between player sales.
Watch: shop unlocks · price recoveryGrow or preserve cash?
Land, animals, seeds, and hands raise future output while reducing the bank that currently decides the match.
Watch: payback day · bank leadProduce or respond?
Every care, harvest, movement, and placement consumes scarce execution capacity.
Watch: overdue work · route lengthExploit or learn?
A rigid plan is efficient against a familiar bot; an adaptive plan can react when the opponent commits elsewhere.
Watch: archetype confidence · switching costReinvest or liquidate?
Late assets only matter if they pay back and their output reaches the bank before turn 720.
Watch: turns remaining · unsold unitsGame-specific claims are grounded in the official Kaggriculture environment reference ↗.
Agent system A
Opponent model
Turn visible commitments into a small, revisable hypothesis.
Inference loop
Observe → classify → respond → update
Never let a weak guess override farm survival.
- 01Observe commitmentsraw signals
Land unlocked, structures built, crop ages, animal mix, visible workers, bank, and shared price movement.
- 02Score candidate archetypesprobability + confidence
Expansion rush, livestock compound, recurring crops, short-cycle cash, or mixed/unknown.
- 03Estimate their next constraintforecast
Cash shortage, feed demand, care load, harvest wave, storage pressure, or liquidation deadline.
- 04Select a bounded responsepolicy modifier
Differentiate the portfolio, preserve liquidity, accelerate a sale, or stay on the baseline when evidence is weak.
- 05Update after evidencecalibration
Decay stale beliefs and compare predicted farm changes with what actually appeared.
Reading windows
What becomes legible when
First seeds, worker count, and early spending reveal tempo but not the full portfolio.
Land and structures make the strategy more expensive to reverse.
Yield cadence, feed load, and price impact expose the bot’s operating model.
New purchases, unsold output, and bank conversion reveal terminal mistakes.
Visible
Opponent farm state and the shared market.
Hidden
Opponent shed, private inventory, source code, and future actions.
Output
Archetype probabilities, predicted constraint, confidence, and response budget.
Agent system C
Lookahead simulator
Chess-like response planning with stochastic branches and partial information.
Rolling horizon
Re-plan every turn; look ahead a few days
Long enough to see payback, short enough to remain responsive.
- STATECompress the boardsₜ
Bank, assets, care deadlines, shed, routes, prices, town demand, opponent hypothesis, and turns remaining.
- OPTIONSGenerate feasible plans3–8 plans
Baseline continuation plus a small set of meaningful deviations: expand, diversify, accelerate, defend, or liquidate.
- BRANCHSample uncertaintymany futures
Opponent response, weed rolls, future shop unlocks, price movement, and execution slippage.
- SCOREEstimate finish outcomesP(bank₇₂₀ > opponent)
Prioritize probability of finishing ahead; use expected bank and failure risk as diagnostics.
- ACTExecute one safe stepreceding horizon
Take the first action from the best robust plan, observe the new state, and simulate again.
Build order
Simulator maturity
Re-run known seeds and reproduce bank, farm, and market transitions.
Starter, expansion, livestock, recurring-crop, and mixed response models.
Sample weeds, shops, and uncertain opponent choices.
Compare predicted win probability with held-out episodes and ladder results.
Agent system D
Attention controller
Protect irreversible obligations before spending effort on adaptation.
Priority stack
What gets attention first
Lower number wins when tasks compete.
- 01Prevent irreversible lossSURVIVAL
Water/feed deadlines, overflow, invalid actions, and products about to decay or cap.
- 02Protect the terminal bankEXIT
As turn 720 approaches, stop purchases that cannot repay and liquidate all reachable value.
- 03Capture time-sensitive valueOPPORTUNITY
Harvest windows, temporary scarcity, shop-driven demand, and care bonuses.
- 04Read the opponentINFERENCE
Update the archetype only when visible evidence materially changes the likely response.
- 05Run lookaheadSIMULATE
Compare the baseline with bounded counter-strategies and choose the robust branch.
- 06Compound the farmGROWTH
Plant, build, hire, expand, and route workers using the remaining action budget.
Guardrails
Adaptation without thrashing
Stay on the baseline until multiple signals support the same opponent model.
A response is only useful if the gain exceeds the cost of abandoning the current plan.
Opponent analysis must not steal in-game actions from mandatory farm work.
Use small reversible deviations to learn; avoid large speculative commitments.
First ladder entry
Ship an interpretable baseline
The first submission is a measuring instrument, not the final strategy.
Learning plan
Submit early, diagnose carefully
- 01Freeze v0.1one known policy
Keep Balanced Tempo readable, deterministic where possible, and protected by a safe fallback action.
- 02Pass local gateszero runtime failures
Complete self-play, starter matches, both seats, varied seeds, and the late-game liquidation test.
- 03Enter the ladderpublic evidence
Kaggle first validates the bot against a copy of itself, then matches it with similarly rated submissions.
- 04Wait for signal50 · 100 · 250 episodes
Record rating, rank, and episode count at consistent checkpoints instead of reacting to every move.
- 05Classify lossesfailure taxonomy
Tag opponent archetype, first strategic divergence, execution failure, and terminal inventory.
- 06Change one mechanismcausal iteration
Improve opponent modeling, lookahead, attention, or a strategy rule—then submit a new version.
Release gate
v0.1 readiness
Self-play3 / 3 seedsboth bots reached DONE
Starter benchmark20 / 20 aheadboth seats · 10 seeds
Runtime failures0across 20 episodes
Terminal shed0–4 items15 / 20 ended empty
Submission
Bot contract
def agent(obs):
return {
"farmer": ["PASS"],
"hands": [],
"market": []
}Entry point
main.py at archive root with agent(obs)ValidationThe first episode runs your agent against a copy of itself
TrackingOnly the latest two submissions remain active for final evaluation
Daily limitUp to five agent submissions per team
Submission and ranking behavior: official Kaggriculture evaluation section ↗.
Reference
Game mechanics
Environment
Fixed variables
The agent observes these rules; it cannot change them.
Match2 players · 720 turns
30 days with 24 turns each.
Starting state3,000 coins · 1 quadrant
Northwest 5 × 5 begins unlocked.
Board10 × 10 tiles
Three more quadrants cost 1,000 / 2,000 / 4,000.
Storage100 shed items
Overflow at day-end is discarded.
Market limit10 orders per turn
Both players affect shared prices.
Basic needs2 missed days
Unwatered plants become weeds; unfed animals escape.
Labor cost1, 1, 2, 3, 5…
Daily hires follow Fibonacci pricing and reset each day.
World noise0.5% weed chance
Shops unlock every 3 days and consume stock every 4 turns.
Board
10 × 10 farm
Economy
Production
| Asset | Buy-in | Base price | First yield | Cadence / life | Daily obligation |
|---|
Prices move with supply and town demand.
Invariants
Never miss
2 daysmissed watering or feed loses the asset
100 itemsshed capacity
Turn 720everything must be sold
Evidence
Experiment log
One change. Enough head-to-head episodes. Clear decision.
Trend
Outcome rate
Log a matchup to start.
History
Runs
No experiments yet
Start with a baseline.
Priorities
Next move
Impact × confidence ÷ effort.
Backlog