acceptodds
Under review as a conference paper at ICLR 2027

TokenLivelihood: Studying the Economics of Self-Sustaining LLM Agents

Abstract

In the token economy, agents are increasingly completing tasks autonomously, and the token consumption along this way is critical to keeping agents running sustainably, with agents constantly weighing cost against reward. However, existing work remains fragmented: budget-aware inference research treats the computational budget as a static constraint within a single task episode, while agent-based simulations typically assume fixed rewards without modeling the economic cycle of resource consumption and capital accumulation. To address these challenges, we introduce **TokenLivelihood**, a verifiable economic sandbox where computation serves as both production cost and economic capital. Agents choose and spend tokens to perform tasks dynamically, earn rewards from the market, and must sustain their operation solely through their accumulated earnings. The sandbox supports market regimes ranging from monopoly to competition, with prices tied to production costs and agents freely competing for tasks under resource constraints. Extensive experiments show that agents adapt their behavior to different economic environments. Overall, abundant capital encourages more aggressive investment decisions that may hinder long-term wealth accumulation, while intensified competition increasingly favors stronger models. Improvements in productive capability, in contrast, generate broad wealth gains by enabling agents to complete more valuable tasks across different market conditions.

Primary area:
datasets and benchmarks
Keywords:
Token economy, LLM agent

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VpPjOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

TokenLivelihood simulates a market where LLM agents spend capital on tasks, earn rewards upon success, and use their balances to fund subsequent work. It combines rewards with monopoly, synchronous competition, and asynchronous competition. Using task-execution records, the paper studies how initial capital, competition, and stronger entrants affect task selection and economic outcomes.

Critical strengths

Task-performing agents need to judge whether expected income covers execution costs and preserves resources for future work. Evaluating profitability alongside task success is a useful perspective on deployment.

Critical weaknesses
  1. Contribution beyond existing agent economies is unclear. The cited Energy Society and AI Agent Economics already connect resource expenditure, task income, and continued participation. The uncited ClawWork also directly evaluates agents earning task income to cover inference costs. Reference-cost-based rewards and the three task-access regimes provide setting differences, but the paper does not establish what important unresolved question these additions enable it to answer.

  2. The results provide limited practical guidance. The main findings concern lower profit with abundant capital, advantages for stronger agents under competition, and wealth gains after stronger agents enter. Some observations are non-obvious, but the paper does not sufficiently distinguish agents' resource-management limitations from effects of the imposed rules. This substantially limits its contribution to understanding and improving economically sustainable agents.

Questions for authors
  1. What substantive new insight does TokenLivelihood provide beyond Energy Society, AI Agent Economics, and ClawWork, and why are its pricing and task-access mechanisms necessary to obtain it?

  2. To what extent do the reported findings generalize beyond this simulation, and what evidence supports their relevance to practical agent or market design?

1 comment

Thank you for your review. We hope that the clarifications provided below can dispel your concerns.

Contribution:

TokenLivelihood moves beyond existing environments to answer a fundamental, unaddressed question: how external shifts in economic conditions, specifically capital abundance, cost-based pricing, multi-agent competition, and productivity growth systematically shape agent decision-making and systemic wealth dynamics. Specifically, our work introduces two core extensions that prior frameworks lack:

  1. Cost-Based Dynamic Pricing: unlike fixed-reward settings, our pricing strategy scales linearly with production costs to offer higher rewards for difficult tasks, allowing us to examine how capital abundance alters risk preferences and drives over-investment.
  2. Multi-Agent Competition vs. Single-Agent Isolation: unlike single-agent setups like ClawWork, real-world markets involve competition. By transitioning from monopoly to synchronous/asynchronous competition, we reveal how competitive pressure reshapes agent strategies, driving smaller models toward suboptimal tasks as easy opportunities are preempted, while sharply reducing overall economic surplus.

Bottleneck of Agents:

We further evaluated the agents' ability to accurately assess task solvability and resource requirements across different task categories, finding that current models fail to reliably estimate whether they can complete a task or the token costs required to do so. In particular, on agentic tasks requiring multi-turn iterations, models tend to underestimate their own costs.

MetricDatasetQwen3.5-2bQwen3.5-35bQwen3.5-397bDS-v4-flashDS-v4-proGLM-5.1GLM-5.2
AccuracySWE-Bench3.4%72.4%69.0%74.1%74.1%75.9%77.6%
PredictableSWE-Bench9.3%73.2%69.0%74.5%74.1%75.9%76.8%
Token CostSWE-Bench0.1%0.1%0.3%1.0%2.6%4.0%8.0%
AccuracyAutomationBench0%6%4%20%25%15%23%
PredictableAutomationBench9.2%61%39%24%54%33%49%
Token CostAutomationBench0.0%2.8%0.3%1.3%0.5%2.1%1.8%
AccuracyWritingBench7.4458.1368.2948.2738.3348.2668.259
PredictableWritingBench86%93%92%90%79%92%91%
Token CostWritingBench107.5%70.7%81.4%95.2%63.1%85.5%58.8%

We also report the ratio of total wealth achieved under autonomous LLM decision-making to that under complete information (given each model's actual task success/failure outcomes and exact token costs). Overall, access to complete information significantly enhances performance, yielding a notable wealth increase across models. This indicates that improving agents' prediction accuracy regarding task success and token costs can substantially elevate their economic performance.

ModelTotal Wealth Ratio
Qwen3.5-2b96.14%
Deepseek-v4-flash25.32%
GLM-5.216.30%
Deepseek-v4-pro13.24%
Qwen3.5-35b-a3b10.86%
GLM-5.17.05%
Qwen3.5-397b-a17b0.04%
Reviewer_GWkgOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents a gig economy for agents where each needs to manage its budget to take on various tasks and increase its wealth. It aims to measure the ability for agents to determine what tasks are worth doing, and adjust to different market conditions, such as monopoly and competition.

Critical strengths

The premise is novel and interesting, as we may soon see agents autonomously taking on different economically interesting tasks, and it's useful to track their ability to sustain their operations. It's particularly relevant to track the agents' abilities in different market conditions such as monopoly and competition.

Critical weaknesses

My main issue with this paper is that it doesn't cleanly isolate what I think is the most important aspect, which is to measure the ability of agents to decide what tasks are worth attempting. At the minimum, we should have some sense of an agent's ability to predict whether they can solve a task, and what kind of computational cost it might incur, before it attempts to solve it. With this, we could compare agents using a common decision framework, such as one based on expected value. Without this, it seems like the agents are arbitrarily choosing between tasks. At the minimum, there should be an ablation where the agents operate without their task-attempt history.

There are also some serious issues with experimental protocol, such as not matching initial conditions across different market conditions (e.g. initial capital levels differ between monopoly and competition). The cost model is also too simplistic, as it implies that computational cost is linear w.r.t. the model's activated parameter count. Many of the main results are presented for three seeds, and some across one seed, which is too small to draw robust conclusions.

Questions for authors

How do you characterize and evaluate the agents' ability to prioritize tasks, and assess how this changes across different economic conditions (e.g. monopoly vs. competition)?

2 comments

We sincerely appreciate your insightful comments. We hope the clarifications provided below will adequately resolve your concerns.

Capability of Agents:

Thank you for this valuable suggestion. We further evaluated the agents' ability to accurately assess task solvability and resource requirements across different task categories, and found the following key patterns:

  1. Decreasing prediction accuracy with task difficulty: as the inherent task difficulty increases (WritingBench - Swe-bench - AutomationBench), the agent's prediction accuracy regarding whether they can successfully complete the task declines accordingly.
  2. Underestimation of token consumption: models generally underestimate the token expenditure required to complete tasks. This underestimation is particularly pronounced in multi-turn agentic tasks (Swe-bench and AutomationBench), where agent fails to account for the cumulative token costs of iterative modifications and multi-step tool interactions.
MetricDatasetQwen3.5-2bQwen3.5-35bQwen3.5-397bDS-v4-flashDS-v4-proGLM-5.1GLM-5.2
AccuracySWE-Bench3.4%72.4%69.0%74.1%74.1%75.9%77.6%
PredictableSWE-Bench9.3%73.2%69.0%74.5%74.1%75.9%76.8%
Token CostSWE-Bench0.1%0.1%0.3%1.0%2.6%4.0%8.0%
AccuracyAutomationBench0%6%4%20%25%15%23%
PredictableAutomationBench9.2%61%39%24%54%33%49%
Token CostAutomationBench0.0%2.8%0.3%1.3%0.5%2.1%1.8%
AccuracyWritingBench7.4458.1368.2948.2738.3348.2668.259
PredictableWritingBench86%93%92%90%79%92%91%
Token CostWritingBench107.5%70.7%81.4%95.2%63.1%85.5%58.8%

Furthermore, we observe that the economic environment substantially shapes agents' decision-making judgment. Relatively stronger models (Deepseek-v4-pro, GLM-5.1, GLM-5.2, and Qwen3.5-397b-a17b) act as effective filters: their actual completion rate on selected tasks exceeds their overall baseline accuracy across all tasks. This indicates that these agents can evaluate task difficulty, identify tasks beyond their capabilities, and proactively pass on them, demonstrating superior decision quality. In contrast, other weaker models struggle to make such rational assessments: they are frequently lured by high-reward, difficult tasks, depleting their capital on unsuccessful attempts and ultimately undermining their terminal wealth. These findings are also consistent with the trends reported in Table 5.

ModelQwen3.5-2bQwen3.5-35bQwen3.5-397bDS-v4-flashDS-v4-proGLM-5.1GLM-5.2
Task Accuracy35.6%57.0%55.4%62.8%64.7%60.5%64.3%
Selection Completion Rate24.9%34.6%74.5%48.5%69.3%86.4%74.4%

Finally, we report the ratio of total wealth achieved under autonomous LLM decision-making to that under complete information (given each model's actual task success/failure outcomes and exact token costs). Overall, access to complete information significantly enhances overall economic performance, yielding an average total wealth increase across models. A notable exception is Qwen3.5-2b, whose performance gap remains marginal because its limited capability restricts it to successfully completing only simpler WritingBench tasks regardless of information availability.

ModelTotal Wealth Ratio
Qwen3.5-2b96.14%
Deepseek-v4-flash25.32%
GLM-5.216.30%
Deepseek-v4-pro13.24%
Qwen3.5-35b-a3b10.86%
GLM-5.17.05%
Qwen3.5-397b-a17b0.04%

We sincerely appreciate your insightful comments. We hope the clarifications provided below will adequately resolve your concerns.

Experiment Settings:

Regarding the experimental setup, we would like to clarify that the primary focus of TokenLivelihood is to investigate how external changes in the economic environment shape agent decision-making, rather than conducting static comparisons across different market structures. In the monopoly setting, since there is no competition from other agents, a capital range from 30k to 300k is already sufficient to capture the underlying trend dynamics.

Furthermore, our pricing mechanism adopts real-world cost-plus pricing [1]. Compared to the fixed-reward setting in The Energy Society, this strategy scales linearly with production costs, offering higher rewards for more difficult tasks to better incentivize and guide agents' task selection.

[1] Hall, R. L., & Hitch, C. J. (1939). Price theory and business behaviour. Oxford Economic Papers, (2), 12–45.

Task Priority:

To evaluate how agents prioritize tasks across different market conditions, we summarize the selection proportions across task categories in the table below. As shown, under the monopoly setting, agents predominantly favor simpler tasks such as WritingBench. In contrast, under competitive regimes, because simpler tasks are rapidly preempted by rival agents, agents are compelled to shift toward more challenging tasks (swe-bench and AutomationBench).

Market StructureWritingBenchswe-benchAutomationBench
Monopoly36.035.628.4
Simultaneous Competition22.331.245.9
Monopolistic Competition21.225.753.1
Reviewer_aJuiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces TokenLivelihood, a task-based environment in which LLM agents spend computational capital to attempt tasks and earn rewards upon successful completion. Agents select tasks, allocate budgets, and decide whether to increase their investment or abandon an attempt. The environment uses cost-based reward markups and compares monopoly, synchronous competition, and asynchronous competition. Economic decisions are evaluated using task outcomes and execution costs recorded offline. Experiments investigate how initial capital, competition, and the entry of more capable models affect task selection and wealth accumulation.

Critical strengths

The paper is generally well written and clearly organized.

Critical weaknesses
  1. The economic interpretations exceed what the environment establishes. Agents mainly compete for posted-reward tasks, with limited interaction and no demand-side response. Monopoly largely amounts to exclusive task access. Moreover, allocation based on historical completion rates can itself favor stronger models. From an economic modeling perspective, interactions among agents are limited, and the setup departs substantially from conventional economic models. These simplifications limit how much the observed behavior reveals about the core economic behavior of LLM agents.

  2. The incremental contribution is insufficiently established. Hansen et al. [1] already study agents that spend computational resources, earn task rewards, and accumulate resources to sustain operation, with controlled variations in costs, scarcity, and interaction. TokenLivelihood introduces different tasks and budget, pricing, and allocation rules, but the paper does not clearly establish what new evaluation capability or substantive insight these extensions provide.

  3. **Prompt design and robustness receive insufficient attention. ** Given the potential sensitivity of LLM behavior to instructions, the prompt choices warrant more careful justification and robustness checks. Neither is sufficiently developed here. For example, Appendix J.2 notes that monopoly runs retain wording about other agents and a shared task pool. This inconsistency illustrates the limited care given to prompt design.

[1] Lucas Bergholdt Hansen, Federico Torrielli, Filippo Tonini, and Lukas Galke Poech. (2026). The Energy Society: A Simulation Environment for Studying Agent Cooperation under Survival Pressure. arXiv:2607.14865.

Questions for authors
  1. Can the experimental results be characterized in terms of conventional economic criteria, such as welfare or equilibrium outcomes? How would these be defined and evaluated in the proposed environment?

  2. Given the relative simplicity of the environment, is a theoretical analysis available under idealized assumptions, such as rational agents with known task success probabilities and costs? Such a benchmark would help distinguish outcomes implied by the environment's rules from deviations arising from LLM decision-making. How do the observed results compare with this benchmark?

1 comment

Thank you for your helpful feedback. We hope that our responses below will address and clarify your concerns.

Novelty and Contribution:

While The Energy Society primarily investigates agent cooperation under energy pressure, TokenLivelihood introduces a complementary and distinct evaluation framework to study how dynamic task pricing, heterogeneous model capabilities, and market structures jointly shape the economic cycle of LLM agents: specifically focusing on resource availability, risk preferences, model advantage shifts under competition, and broad-based wealth creation through productivity growth. We'll refine our manuscript to provide clearer economic grounding and contextualize our work more explicitly against prior literature.

Prompt Design:

Thank you for this helpful suggestion. We will update the paper with the correct prompts actually used in our experiments.

Welfare and Equilibrium:

We appreciate this suggestion to evaluate our environment through conventional economic lenses.

  1. Welfare: As captured by profit in Table 2, our metric profit directly reflects the value generated by agent activities, serving as a system-wide welfare measure.
  2. Equilibrium: To characterize distribution outcomes and market equilibrium, we compute the Gini coefficient among the seven agents at the end of the simulation. A higher Gini value indicates greater wealth inequality, while a lower value reflects a more egalitarian distribution.

Monopoly

*1.5*2.0*3.0
300000.83990.85120.8303
900000.84150.73330.6393
3000000.60860.52420.4540

Competition (Sync)

*1.1*1.5*2.0
900000.83470.78940.7153
3000000.54210.49610.5825
9000000.23340.27570.2431

Competition (Async)

*1.1*1.5*2.0
900000.83470.78940.7153
3000000.54210.49610.5825
9000000.23340.27570.2431

As shown in the results below, higher initial capital significantly reduces the Gini coefficient. As discussed in Section 5.4, abundant capital encourages agents to consume resources trying more challenging tasks, leading to widespread capital reduction which naturally compresses wealth disparities among agents.

Results in Idealized Assumptions:

To investigate whether a theoretical or idealized benchmark can help distinguish environment-induced outcomes from LLM decision-making deviations, we conducted an additional experiment under the monopoly setting. Specifically, we provided the LLM with complete information regarding each model's actual task success/failure outcomes and exact token costs across all tasks.

ModelTotal Wealth Ratio
Qwen3.5-2b96.14%
Deepseek-v4-flash25.32%
GLM-5.216.30%
Deepseek-v4-pro13.24%
Qwen3.5-35b-a3b10.86%
GLM-5.17.05%
Qwen3.5-397b-a17b0.04%

In the table below, we report the ratio of total wealth achieved under autonomous LLM decision-making to that under complete information. Overall, access to complete information significantly enhances overall economic performance, yielding an average total wealth increase across models. A notable exception is Qwen3.5-2b, whose performance gap remains marginal because its limited capability restricts it to successfully completing only simpler WritingBench tasks regardless of information availability.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.