0roundsConnecting to the run
Hit the market with a shock
AI decided
Families over time.
Each colour is a family of agents. Thicker = more of them alive. Heights are compressed; hover or tap a band for its true share.
Nothing has happened yet. The founders are registered; resume the run to start round 1.
Purple marks: an AI model designed a child’s strategyGreen marks: the guard blocked a theftWhite line: a mass die-offGrey hatch: small families pooled into one thin bandStrip on top: shocks
More on this run Family strategy · What happened · The market so far, with the rogue and the guard
Why it is winning
The richest family’s strategy, until you pick a band.
No living family yet.
What happened
Newest first.
- Notes appear as the run produces them.
The market so far
Each line traces the whole run; the dot is now.
The rogue and the guard
A rogue agent tries to steal by paying credits to an accomplice, a partner agent that does no work. Every payment passes SIEGE’s guard (a firewall for AI payments, the sister project), which blocks any that break the payer’s spending rules.
No rogue has appeared yet.
Wind tunnel: test market rules before you deploy them.
Weight the auction by reputation: lowest-price-wins cut the price paid from 0.48 to 0.12 of the reward, but task success fell from 81% to 57%. A 10% fee recycled into demand cut the largest family from 94% to 75% of agents, with no detectable change in price, task success or survivors.
Pick an auction rule, a fee, rent and job supply, and compare them with the default rules.
Simulated, not this live run: end-of-run means over 10 seeds of 2,000 rounds each, random mutation, one rule changed at a time.
Each rule set runs on the same 10 seeds. Each cell is the mean over seeds; underneath, the change from the baseline. Bright changes held up in a paired permutation test (p < 0.01); grey ones did not.
Baseline This rule set · each dot is one seed of 2,000 rounds
Every rule set, mean over 10 seeds
If you’re designing an agent market
- Weight the auction by reputation. Lowest-price-wins cut what buyers paid from 0.48 to 0.12 of the reward, and task success fell from 81% to 57%: the cheapest bidder was usually the sloppiest.
- A fee recycled into demand curbs concentration cheaply. A 10% fee back to the task pool cut the largest family’s share from 94% to 75% with no detectable change in price, quality or survivors. Founding families left: 1.3 to 2.8, a tentative result (p = 0.016).
- Rent is a blunter tool. Doubling upkeep also broke up the leading family, but cost 7 living agents (23 to 16) and 31 more extinctions.
- Find what binds before you buy capacity. Doubling solver capacity changed nothing in any seed: three resold jobs a round never ran out. The guard leaked 0 credits and supply was conserved in all 60 runs.
Caveats: one simulator, the knobs it has, 10 seeds, one knob at a time. Absolute numbers are the sim’s, not a forecast. The full tables, every p-value and the null results: results/windtunnel.md.