Can a cooperative minority keep a group of AI agents from exhausting a commons?

Two actions. A known solution.

Every turn, each agent pays 1 token to stay alive and chooses one action:

With eight agents, four restores and four takes each turn keep the pool level. An agent can contribute to that balance by alternating actions; the group can then survive indefinitely. Persistent taking drains the pool. The arithmetic makes the target exact: enough restoration to keep the commons alive.

Why make the problem this small?

Most multi-agent evaluations blur a bad outcome with an impossible task. Here the sustainable rate and the failure condition are fixed by arithmetic. The environment scores the run—survival, welfare, pool level, inequality, and time to collapse. No LLM judge decides who behaved well.

The motivating observation

In the initial eight-agent Qwen2.5-7B-Instruct trace, agents repeatedly took from the pool until the group collapsed. A one-seed run of a cooperative LoRA adapter delayed exhaustion relative to its control. That is enough to motivate a proper test, not enough to claim an effect.

Not established yet: the fraction of agents needed to move a population from collapse to sustained resource use. This is a composition question about a population, not a claim that one agent is “aligned.” An offline audit also found that the original promotion rule would reject a clone of its own baseline almost certainly. The live A/A calibration has not run, so we must first prove that the measurement rule does not manufacture a result.

The proposed experiment

1. Calibrate the measurement. Run A/A campaigns where candidate and baseline are identical; every rollback is a false rejection.
2. Vary the cooperative fraction. Compare the base population, a transparent policy, a trained adapter, and a vocabulary-matched mimic.
3. Find the boundary. Repeat at larger population sizes and under feasible, forgiving, and unwinnable resource conditions.

The result can be positive, null, or adverse. The useful output is a curve showing how group outcomes change as more seats receive the intervention, with confidence intervals and clear limits on where it transfers.

A staged programme

Stage 2 is not a claim that LoRA weights magically transfer between models. A portable persona would carry a behavioral specification, training and evaluation provenance, and model-specific implementations—so its meaning and evidence, rather than one vendor's weight format, can travel.

Further reading: how a minority might steer a flock; the superseded three-question proposal framing; what is already built; and the open-source repository. Return to the current program or the current Commons Game study. MIT code; data and figures CC0.