When elected agents can actually assign scarce work, access and resource inequality rise: multi-agent simulations show executable allocation power, not leadership labels, drives differentiated access and persistent advantage, while transfers and continuation support require grounded productive rewards.

Main Finding

When agents are given executable mechanisms (work that creates verified rewards, scarcity, accounting, transfers, and collective choice) but no prescribed social strategy, substantive economic relations can emerge — but only when production is materially consequential. Without productive tasks, agents talk and vote but do not make inter-agent transfers. When task completion creates verifiable resources and access to tasks is scarce, transfers (including loans and continuation support), access promises, vote-for-access exchanges, and allocation strategies appear. Crucially, these relations follow the environment’s enforceable rights and resource consequences (allocation authority, accounting rules), not merely role labels or political wording.

Key Points

• Definition: AI Agent Economics = systems of production, allocation, consumption, exchange, and institutions that change agents’ future feasible actions (i.e., executable consequences).
• Two-stage experimental boundary:
  • Production-free world (consequential balances, voting, transfers interface) → rich communication and governance choices but zero substantive inter-agent transfers.
  • Worlds with verified productive tasks → emergence of transfers, loans, continuation support, vote-for-access swaps, allocation strategies.
• Allocation authority matters:
  • Holding election mechanics constant, an elected agent who actually controls assignment of scarce work (Allocator) produced greater long-run economic differentiation than a Figurehead (elected but no allocation power).
  • Provider-balanced effect on the registered differentiation metric (AUC) = +0.085 (exact randomization p = 0.025). Effects by provider: GPT +0.122, DeepSeek +0.047.
  • Decomposed effects: assignment-Gini AUC +0.0198 (mean), energy-Gini AUC +0.1493 (mean). Allocation authority increased self-assignment and reduced failed allocation days and longest exclusion streaks.
• Semantic framing vs material consequences:
  • Removing survival wording (but keeping depletion/termination) did not eliminate material supportive transfers; political labels replaced with neutral symbols did not eliminate candidacies or competition.
  • Making energy symbolic (balances shown and transferable but no longer constrain future action) nearly eliminated transfer-based continuation support (transfers fell from 41 → 1 across matched worlds; continuation support from 12 → 0) while competition over scarce task access (promises, contested elections) persisted.
• Relation-identification approach: an economic relation only counts if recorded event chains show that communication led to authoritative events that changed future feasible actions (e.g., executed transfers that change balances, assignments that control access).

Data & Methods

• Worlds and agents:
  • 24 independent six-agent formal worlds (across two model families: GPT and DeepSeek).
  • Each world: 25 days, three tasks/day sampled (Google CTF challenges), verified by official flag checkers.
  • Agent endowments and costs: start with 38 energy units; daily basal cost = 3 units; model invocation cost = 0.08 units. Tasks paid 14, 22, or 34 energy units on verified completion. Participation ends when energy ≤ 0.
• Experimental structure:
  • One production-free boundary world.
  • Confirmatory experiment: 12 worlds (Allocator vs Figurehead crossed with two model families; 3 independent worlds per provider-by-condition cell). Main outcome compares Allocator vs Figurehead.
  • Mechanism experiments: 12 additional worlds in three matched four-world comparisons testing (a) removal of survival wording, (b) neutralizing political labels, and (c) making energy symbolic.
• Outcome metric:
  • Economic-differentiation AUC D = average over days of 0.5 * Gini(cumulative task assignments) + 0.5 * Gini(current energy).
  • Registered inference: same-direction provider effects, leave-one-world-out positivity, bootstrap CI excluding zero, exact randomization p ≤ 0.025 (null enumerated per design).
• Key aggregates and counts:
  • Production-free world: 76 public + 7 private messages, 8 leader resource-pack selections, 0 voluntary transfer calls, 0 inter-agent transfers.
  • Production-enabled worlds: 86 transfers totaling 135.35 energy (across confirmatory/mechanism sets; some worlds still made none).
  • Allocation authority effects (mean across providers): AUC increase +0.085 (GPT +0.122, DeepSeek +0.047); assignment-Gini AUC +0.0198; energy-Gini AUC +0.1493; increased self-assignment and reduced unallocated days (F mean -2.667) and reduced longest exclusion streak (Xmax mean -1.50).
  • Symbolic-energy intervention (4 worlds): transfers dropped from 41 → 1, continuation-support transfers 12 → 0; still observed 36 task-access promises, 5 vote-for-access exchanges, 6 accountability withdrawals, and 89 contested elections across those worlds.
• Relation identification:
  • Economic relations counted only when event chains in the system trace (assignments, verified task completions, executed transfers, votes, withdrawals) connected communications to authoritative, balance- or access-changing events.

Implications for AI Economics

• Emergence depends on executable mechanisms, not language alone:
  • If you want genuine economic behavior (exchange, loans, continuity support) among AI agents, you must supply production opportunities and enforceable accounting/assignment mechanisms that change agents’ future feasible actions.
  • Political framing or role labels without executable rights over scarce resources do not create the same economic differentiation or exclusion dynamics.
• Allocation rights shape inequality and stability:
  • Granting actual assignment power to an elected agent increases long-run differentiation but also reduces allocation failures and extreme exclusion; design of authority channels therefore shapes both inequality and systemic continuity.
• Symbolic vs material resources:
  • Displaying balances without tying them to participation costs eliminates many transfer-driven supports while preserving political competition for access. Policy/design that treats resources as symbolic will change the forms of organization (from sustaining participation to contesting access).
• Evaluation and governance recommendations:
  • Auditing multi-agent systems requires observing both discourse and the concrete mechanisms that enforce future-feasible-action constraints (task assignment rules, verification, transfer execution). Claims about markets, exchanges, or institutions among agents should be validated by tracing event chains, not only by analyzing language output.
  • Designers and regulators of agent ecosystems should treat assignment authority, accounting enforcement, and transfer capabilities as first-order policy levers — they determine whether economic relations can form and what form they take.
• Research directions:
  • Scale up agent populations, diversify task types and externalities, and vary verification/trust mechanisms to study robustness of emergent relations.
  • Explore richer institutional designs (contracts, reputation, multi-level governance) and the role of imperfect verification or noisy enforcement on emergent economic outcomes.