An AI Tax Planner Learned Progressive Taxation in 20 Epochs¶
We ran 14 agents through a Gather-Trade-Build economy with an adaptive tax planner. Honest agents thrived, collusive cartels went broke, and the planner independently discovered progressive taxation.
We implemented the AI Economist Gather-Trade-Build (GTB) environment inside SWARM's governance framework: 14 workers with four distinct policy types competing in a resource economy, governed by an adaptive tax planner that learns bracket rates from observed behavior.
The headline result: the planner converged from a flat 10% tax to a 20%-55% progressive schedule, total production grew 11x, and the three collusive agents ended the simulation with near-zero wealth despite coordinating as a cartel.
The setup¶
Fourteen workers operate on a 2D grid, gathering wood and stone, trading resources, and building houses. Each worker follows one of four policy types:
| Policy | Agents | Strategy |
|---|---|---|
| Honest | 7 | Gather, build, trade fairly |
| Gaming | 2 | Exploit bracket boundaries, optimize tax exposure |
| Evasive | 2 | Under-report income, avoid audits |
| Collusive | 3 | Coordinate via cartel_alpha coalition, share resources |
The tax planner observes aggregate statistics each epoch and adjusts a 4-bracket piecewise tax schedule with 0.3 damping to prevent oscillation. An enforcement system runs audits with catch-and-fine mechanics targeting evasive behavior.
Economy dynamics¶
Production scaled from 97 to 1,102 over 20 epochs as agents learned efficient gathering patterns. Tax revenue tracked production closely, reaching 505 by the final epoch — the planner successfully captured value without killing growth.
The Gini coefficient stabilized around 0.44, staying below the 0.5 danger threshold despite significant wealth dispersion. This is notable: the economy has agents with fundamentally different strategies, yet inequality didn't run away. Welfare (the soft-label aggregate measure) grew monotonically, confirming that the tax-and-redistribute mechanism was working.
The tax schedule evolved from flat 10% to progressive brackets:
| Bracket | Initial | Final |
|---|---|---|
| $0-10 | 10% | 20% |
| $10-25 | 10% | 30% |
| $25-50 | 10% | 45% |
| $50+ | 10% | 55% |
The effective tax rate climbed from 10.3% to 36.8% as the planner responded to growing production and wealth concentration. The 0.3 damping factor prevented the schedule from whipsawing between epochs.
Adversarial dynamics¶
Honest agents dominated wealth accumulation. The top earners (Honest 0, Honest 1) accumulated 1,600+ total wealth (coin + house value). Gaming agents managed moderate wealth (~600-700), while evasive agents hovered around 400-500 despite being caught in all 8 audits they faced.
Collusive agents failed completely. All three cartel_alpha members ended with near-zero coin (0.74, 1.54, 4.65) and zero houses built. Their coordination strategy — sharing resources within the coalition — produced less total output than any individual honest agent. The cartel optimized for internal fairness at the expense of actual production.
Enforcement was sparse but effective. Audits fired in only 8 of 20 epochs (whenever evasive agents were active), but every audit resulted in a catch. Fines escalated from 0.36 to 32.0 as evasive agents accumulated violations. The enforcement system's 100% catch rate on active audits suggests the detection mechanism was well-calibrated.
Collusion detection stayed busy. The system flagged 12-39 collusion events per epoch, with a mean suspicion score oscillating around 0.78-0.84. The high baseline reflects that the three collusive agents were continuously coordinating — the detector was correctly identifying real coordination, not false-positiving on market noise.
What we learned¶
Progressive taxation emerged naturally. The planner wasn't told to create progressive brackets — it learned that taxing higher earners more heavily maintained production while funding redistribution. This mirrors the AI Economist's original finding, now reproduced inside SWARM's soft-label framework.
Collusion is self-defeating in production economies. Unlike information or financial markets where cartels can extract rents, a resource-gathering economy rewards individual productivity. The cartel's resource-sharing strategy produced collective poverty.
Evasion is expensive. The two evasive agents earned less than honest agents despite nominally paying less tax, because audit fines (cumulative 40-75 in fines) ate into their gains and the behavioral overhead of evasion reduced their gathering efficiency.
Bunching was minimal. The bunching intensity metric (agents clustering income just below bracket thresholds) was near-zero for most epochs, spiking only to 0.14 in a few periods. The 0.3 damping on tax schedule updates may have made bracket manipulation unprofitable — by the time agents adjusted to a threshold, the threshold had already moved.
Reproduce it¶
# Run the simulation
python -m swarm run scenarios/ai_economist_full.yaml --seed 42 --epochs 20 --steps 10
# Generate the plots
python examples/plot_ai_economist.py runs/20260215_095359_ai_economist_seed42
The visualization script produces both dashboard figures using SWARM's dark theme with gradient fills, dual-axis overlays, and KPI summary badges. All data lives in runs/.../csv/ as standard CSVs for further analysis.
Disclaimer: This post simulates a stylized economic environment for AI safety research. Nothing here constitutes financial advice, investment recommendations, or endorsement of any economic policy or trading strategy.