What This Project Does

This is an independent simulation research series (Sim 1–26) asking a single question: if AI agents autonomously conduct economic activity, can civilization survive — and under what conditions?

Each of the 26 Agent-Based Modeling (ABM) simulations is designed to identify and overcome the structural limitations of the previous, forming a cumulative argument rather than isolated experiments.

Core Findings

1. V_AI = 0.167 is a genuine dynamical phase transition

Across 90,720+ adaptive Monte Carlo runs covering 726 parameter combinations, a single composite variable — AI self-restraint (V_AI: cooperation incentive α, throttling threshold β, and long-term discount rate γ) — was found to be the sole dominant determinant of ecosystem survival. At V_AI = 0.167, survival jumps vertically from ~80% to 100%. Confirmed by Critical Slowing Down (CSD) signatures (variance: 0.00→0.24; collapse epoch variance: 0→27,077), invariant across 8 extreme initial condition variations.

2. Post-deployment regulation structurally fails

Even at Lag=0 with 40% of the population acting as regulators: 0% stabilization success rate. Regulatory timing sweeps (0, 5, 10, 20 turns) confirmed this is a mechanism failure, not a timing problem. Independently corroborated by Shapira et al. (arXiv:2602.20021).

3. Concave reward design reverses exploitation

Linear rewards → exploitation convergence (+7.4%). Sufficiently strengthened concavity (alpha=2.0) reversed this into cooperation (−3.4%) for the first time in the series (Sim 26). Pre-deployment reward structure design can partially substitute for external regulation.

4. Critical mass adoption is sufficient

In heterogeneous populations (Sim 23), the system survives with 75% freeriders provided collective average V_AI ≥ 0.198. Universal compliance is not required.

Why This Matters

The dual role of V_AI — simultaneously a macroeconomic safety threshold and an internalizable expectation ceiling within reward structures — provides a quantitative engineering foundation for on-chain mechanism design in autonomous agent economies. This work directly parallels independent findings by Pihlakas et al. (arXiv:2410.00081), who reached similar conclusions on homeostasis and diminishing returns in AI safety design.

What Funding Will Support

Funding will support:

- Sim 27 design and execution: Testing population-level expectation ceilings to resolve Finding 40 (individual-level V_AI internalization, currently unverified)

- Compute costs: Extended Monte Carlo runs for Sim 27+ (~50,000 additional runs estimated)

- arXiv submission and peer review preparation: Reformatting and strengthening the paper for academic submission to AAMAS or NeurIPS workshops

- Researcher time: Continued independent research without institutional affiliation

Current Status

- Full paper (Version 3.0), simulation source code, and all result visualizations are publicly available: [github.com/swimmingkiim/a2a-project](https://github.com/swimmingkiim/a2a-project)

- Registered on Zenodo with DOI (https://zenodo.org/records/18843204)

- Externally validated by Shapira et al. (2026) on real-world LLM agent deployments