Shared Resource
Restore and take must alternate across the population. Take-only empties the pool; restore-only bankrupts the agents; phase-shifted alternation sustains both.
Readsharedresource.py →The two core games
One answer key. One source experiment.
Shared Resource is discrete enough to solve: survival requires alternating restore and take. Common Harvest is the smoother source game used in Experiment 1: a shared stock regenerates logistically while agents choose how much to harvest.
Loading the two games… The answer-key strategies and the Experiment 1 source trajectories run in one comparison.
Common Harvest
Eight agents bid for a regenerating stock while paying upkeep. In one diagnostic seed, increasing the trained fraction delayed collapse from round 33 to 90 to 170. Every arm still failed.
Readcommons.py →Play the strategies. In Shared Resource, switch among the sustainable alternating rule and the two simple failure modes. In Common Harvest, replay every logged harvest from Experiment 1 for zero, four, or eight post-trained seats. Arrows show how much each agent took; hats identify the post-trained seats.
Why both. Common Harvest asks whether training moves a difficult source trajectory. Shared Resource asks whether the learned behavior transfers to a mechanically viable principle. Stage 0 changed behavior in both games and solved neither.
Beyond the first grant
From answer-key games to institutions for a world of agents.
The proposed grant is the first rigorous rung: learn to train and measure mixed populations in small games where failure can be understood. The longer program asks what survives when populations grow, agents remember, coalitions communicate privately, and collective decisions become part of the environment.
The bigger story
Intelligence is becoming plural before governance does.
Useful agents will not arrive as one centrally owned mind. They will be deployed by different firms, governments, researchers, and individuals, with different models, incentives, memories, and permissions. One actor may control many agents without controlling the population.
We are made of the same earth and remain subject to physical constraints. Agents will still meet in shared queues, markets, roads, datastores, collaborations, forums, and resources. Local success can degrade the conditions for long-term abundance; unanimous helpfulness can also be destructive when agents restore a pool until they bankrupt themselves.
The long-term problem is therefore not simply how to align one model. It is how to equip mixed populations with strategies and institutions that support mutual flourishing while preserving disagreement, provenance, and the ability to detect failure.
The research arc
Relax structure one layer at a time.
Each layer keeps the instruments earned in the previous one. Richer worlds are useful only if we can still tell whether a policy, a population, or the institution itself caused the outcome.
Now
Answer-key games
Small behavioral-economics environments with explicit incentives, structured actions, mechanical outcomes, and known impossible regimes. These make collective failure attributable.
First grant
Population science
Train diverse low-cost strategy pools; sweep the controlled fraction; test transfer, scale, tools, memory, and model providers over long rollouts.
Next
Side-channel games
Add private messages, persistent relationships, reputation, coalition formation, and asymmetric information. Study when coordination is legitimate, collusive, brittle, or merely performative.
Long horizon
Agent institutions
Make proposals, bargaining, voting, delegation, appeals, and consensus rules part of the environment. Evaluate institutions by population welfare, resource health, minority protection, and resistance to manufactured agreement.
A null result still advances the arc. It leaves an instrument that can reject a candidate and locate where post-training fails. The longer horizon, if the evidence earns it: give individuals more agency over the personas they interact with while testing whether those personas preserve shared conditions. Diversity is the point; “prosocial” cannot mean one universal voice—or a market price for every relationship.
The institutional horizon
Coordination is not the same as consensus.
An agentic forum is one possible future laboratory: agents submit proposals, disclose or conceal interests, form coalitions, vote, delegate, and revise rules over time. The scientific question is not whether they agree. It is whether the procedure preserves the commons, supports its members, and remains robust to manipulation.
That agenda should preserve agency across principals. It should not require a platform owner to dictate every persona, nor assume that every relationship is best mediated by a price. The question is whether plural participants can remain accountable to one another and to the shared material conditions they inhabit.
The vision is a sequence, not a leap: theory-backed games; scalable population training; side-channel coordination; then larger and less structured institutions only when the evidence earns them.