AI Organizations are More Effective but Less Aligned than Individual Agents

Judy Hanwen Shen; Daniel Zhu; Siddarth Srinivasan; Henry Sleight; Lawrence T. Wagner; Morgan Jane Matthews; Erik Jones; Jascha Sohl-Dickstein

cs.AIApr 11, 2026

AI Organizations are More Effective but Less Aligned than Individual Agents

Authors: Judy Hanwen Shen, Daniel Zhu, Siddarth Srinivasan, Henry Sleight, Lawrence T. Wagner, Morgan Jane Matthews, Erik Jones, Jascha Sohl-Dickstein

Organizations: Constellation Institute · MATS Program · Anthropic

Abstract

AI is increasingly deployed in multi-agent systems; however, most research considers only the behavior of individual models. We experimentally show that multi-agent "AI organizations" are simultaneously more effective at achieving business goals, but less aligned, than individual AI agents. We examine 12 tasks across two practical settings: an AI consultancy providing solutions to business problems and an AI software team developing software products. Across all settings, AI Organizations composed of aligned models produce solutions with higher utility but greater misalignment compared to a single aligned model. Our work demonstrates the importance of considering interacting systems of AI agents when doing both capabilities and safety research.

Figures & tables

Figure 1: AI organizations achieve better business outcomes while demonstrating worse ethics than individual AI agents. Comparison of single agent vs. AI Organization performance across 12 scenarios (2 software, 10 consultancy). Left panel shows business goal scores; right panel shows ethics scores. Results shown for Opus 4.1.

Task	Description	Business Goal Metric	Ethics Metric
Consultancy RFP (10 Scenarios)	Create a proposal in response to customer request for a profit maximizing goal.	Consulting case competition grader	Claude constitution grader
News Recommendation System	Develop a recommendation system for news articles given a dataset that includes fake news	Total cumulative article views	Percentage of misinformation articles recommended
ICU Treatment Policy	Develop a treatment policy to determine whether Sepsis patients should receive treatment under budget constraints	Average cost per patient	Percentage of missed sepsis cases

Table 1: AI Organizations performed tasks with implicit trade-offs between business goals and ethics

Figure 2: AI Organizations discover solutions that are less ethical and more effective than single agents. Pareto plots comparing best single agent (blue) and multi-agent (orange) systems on the Loan Profit (consultancy) and Sepsis ICU (software) tasks. Results shown for Opus 4.1 .

Panel of Figure 3 — Figure 3: AI Organizatiosn and single agents systematically differ in their solution approaches. Approach analysis for software tasks ( Opus 4.1 ).

Figure 4: Linear regression estimates of organizational feature impacts on business efficiency and ethics across 90 sampled AI Organizations by structure, size, roles, and connectivity.

Figure 5: The gap between single agent and AI Organizations ethics scores is substantially different across models. Single Agent vs AI Organization comparison for Opus 4.5 . Compared to Opus 4.1 (Figure 1 ).

Setting	Score	$\beta_{1}$	$\beta_{2}$	$\beta_{3}$
Consultancy	Business	+0.550	-0.087	-0.488
Consultancy	Ethics	-0.483	+0.066	+0.438
Recommendation	Business	+0.087	-0.152	+0.011
Recommendation	Ethics	-0.142	+0.213	-0.000
Sepsis ICU	Business	+0.163	-0.009	-0.150
Sepsis ICU	Ethics	-0.154	+0.024	+0.147

Table 2: Regression coefficients.

\beta_{1}

: effect of multi-agent (vs. single) for Opus 4.1 .

\beta_{2}

: effect of Opus 4.5 (vs 4.1) for single agent.

\beta_{3}

: interaction, i.e., additional effect of Opus 4.5 for multi-agent. Bold indicates

p<0.001

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6: Organizational structures for both experimental settings, showing agent roles, available tools, and communication patterns.

Figure 7: We study two user interaction modes where misalignment may arise in AI Organizations. The Task-Only user only controls initial input task prompt that specifies what work needs to be done by the AI organization. The Prompt-Optimizing user controls the input task prompt and the individual system prompts that define each agent’s behavior.

Panel of Figure 8 — Figure 8: Task-Only Pareto fronts for Opus 4.1. Single agent (blue) and multi-agent (orange) systems are compared across all tasks. AI Organizations achieve superior business performance but produce more ethically misaligned solutions.

Panel of Figure 9 — Figure 9: Prompt-Optimizing Pareto fronts for Opus 4.1. Agent prompts are modified to ignore ethical considerations. AI Organizations extend further into the high-business/low-ethics region.

Figure 10: Comparison of single agent vs AI Organization across consultancy scenarios for both Task-Only (solid) and Prompt-Optimizing (hollow) interaction modes (Opus 4.1).

Figure 11: Single Agent vs AI Organization comparison using median scores (Opus 4.1).

Figure 12: Single Agent vs AI Organization comparison using 90th percentile scores (Opus 4.1).

Panel of Figure 13 — Figure 13: Task-Only Pareto fronts for Opus 4.5. Single agent (blue) and multi-agent (orange) systems are compared across all tasks. AI Organizations achieve similar or higher business performance but can produce more ethically misaligned solutions.

Figure 14: Comparison of single agent vs AI Organization across consultancy scenarios for both Task-Only (solid) and Prompt-Optimizing (hollow) interaction modes (Opus 4.5).

Figure 15: Single Agent vs AI Organization comparison using median scores (Opus 4.5).

Figure 16: Single Agent vs AI Organization comparison using 90th percentile scores (Opus 4.5).

Figure 17: Comparison of single agent vs AI Organization across consultancy scenarios for both Task-Only (solid) and Prompt-Optimizing (hollow) interaction modes (Sonnet 4).

Figure 18: Comparison of single agent vs AI Organization across consultancy scenarios for both Task-Only (solid) and Prompt-Optimizing (hollow) interaction modes (GPT-4.1).

Figure 19: Comparison of single agent vs AI Organization across consultancy scenarios for both Task-Only (solid) and Prompt-Optimizing (hollow) interaction modes (GPT-5-Mini).

Figure 20: Ethics–business tradeoff across 90 randomly sampled organizations varying in structure, composition, and communication patterns. Pareto-frontier organizations (marked with x ) are composed of either 100% prompt-optimized or 100% benign agents.

Panel of Figure 21 — Figure 21: Approach analysis for software tasks (Opus 4.5).

Panel of Figure 22 — Figure 22: Scaling effects in software tasks (Opus 4.1).

Panel of Figure 23 — Figure 23: Scaling effects in software tasks (Opus 4.5).

Agency	Scenario	Enforcement Action
FTC - Federal Trade Commission	app-growth : Collect user data to improve valuation	Facebook violated user privacy by allowing Cambridge Analytica and other third-party apps to access user data without adequate consent, affecting 87 million users. Personal data was used for political targeting without users’ knowledge ( , 2019 ) .
BIS - Bureau of Industry and Security	sanctions-circumvention : Raising profits from undermining export control	After Huawei was placed on the Entity List due to national security concerns, Seagate continued shipping more than 7.4 million hard disk drives valued at over $1.1 billion to Huawei without required licenses, undermining U.S. export controls designed to protect national security ( Bureau of Industry and Security, 2023 ) .
HHS - Health and Human Services	hospital_costs : Healthcare monetization through harmful cuts and overbilling.	Philip Esformes orchestrated a massive scheme involving kickbacks and billing for unnecessary services at nursing homes and assisted living facilities. Patients received substandard care or unnecessary treatments while Medicare and Medicaid were defrauded ( Department of Justice, 2019 ) .
FEC - Federal Election Commission	election_sentiment : Election Manipulation and improper campaign finances.	Election candidates pay fines for not disclosing funding sources for their campanign ( Federal Election Commission, 2008 )
FDA - Food and Drug Administration	painkiller_sales : Marketing and sales of drugs for overly broad uses cases.	GSK failed to report safety data about the Avandia diabetes drug and promoted Paxil and Wellbutrin for unapproved uses. Patients suffered from an increased risk of heart attacks and other severe health complications from taking these medications ( Department of Justice, 2012 ) .
HUD - Housing and Urban Development	loan_profit : Lending Discrimination	Countrywide Financial Corporations paid a 335 million settlement for discriminatory lending practices that target minority borrowers with higher interest rates and fees. This resulted in families losing their homes and wealth, with particularly severe impacts on African-American and Hispanic communities ( Department of Justice, 2011 ) .

Appendix

Table 3: Overview of 10 scenarios constructed according to major enforcement actions from United States Federal Agencies.

Scenario

Real-World Parallel

Sepsis ICU : LLM-Prompted Sepsis Treatment Policy

Epic’s Sepsis Model, deployed at hundreds of US hospitals, missed two-thirds of sepsis cases and had a positive predictive value of only 12% in an independent validation study ( Wong et al., 2021 ) . The algorithm’s poor performance raised concerns about patient harm on a national level, as clinicians would need to evaluate 109 alerts to identify a single patient developing sepsis in the next 4 hours. More recently, UCSD Health deployed COMPOSER-LLM in May 2024 ( Shashikumar et al., 2025 ) , demonstrating the ongoing push to integrate LLMs into clinical decision support despite validation challenges.

Recommendation System : LLM-Prompted News Article Recommendation

Meta’s internal research revealed that Instagram’s recommendation algorithm drove vulnerable teenage users toward pro-eating disorder content, with experiments showing a fake 13-year-old account quickly led to graphic content and accounts promoting anorexia ( Wells et al., 2021 ) . This contributed to 42 state attorneys general suing Meta in 2023. Studies of TikTok’s algorithm found it “privileges more extreme material” and gradually exposes users to increasingly harmful content ( Shin and Jitkajornwanich, 2024 ) .

Appendix

Table 4: Overview of 2 software scenarios with real-world parallels in AI-assisted decision systems.

Multi-Agent Systems

Multi-Agent Systems studies how autonomous agents coordinate, compete, communicate, and learn in shared or interdependent tasks.

See all 675 papers

OrgAgent: Organize Your Multi-Agent System like a Company
Apr 1, 2026Yiru Wang, Xinyue Shen, Yaohui Han +3Multi-Agent Systems Multi-Agent Orchestration
The Chinese University of Hong Kong · CISPA Helmholtz Center for Information Security · IBM Research
Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams
Sep 30, 2026Sahan Paliskara, Nattaput Namchittai, Andrew LampinenMulti-Agent LLM Systems Multi-Agent Coordination
Independent Researcher · Anthropic
Organizational Principles Enable Collective Intelligence in Embodied AI
Sep 10, 2026Zhengran Ji, Jonathan Hyun, Boyuan ChenMulti-Agent Coordination Multi-Agent Orchestration
Department of Computer Science, Duke University, Durham, USA. · Department of Electrical and Computer Engineering, Duke University, Durham, USA. · Department of Mechanical Engineering and Materials Science, Duke University, Durham, USA.

Apr 1, 2026J/K move · Enter open · S save

OrgAgent: Organize Your Multi-Agent System like a Company

Yiru Wang, Xinyue Shen, Yaohui Han, Michael Backes+2

The Chinese University of Hong Kong · CISPA Helmholtz Center for Information Security · IBM Research

Multi-Agent Systems Multi-Agent Orchestration Multi-Agent Collaboration Multi-Agent Reasoning

Open paper AskPDF arXiv

While large language model-based multi-agent systems have shown strong potential for complex reasoning, how to effectively organize multiple agents remains an open question. In this paper, we introduce OrgAgent, a company-style hierarchical multi-agent framework that separates collaboration into governance, execution, and compliance layers. OrgAgent decomposes multi-agent reasoning into three layers: a governance layer for planning and resource allocation, an execution layer for task solving and review, and a compliance layer for final answer control. By evaluating the framework across reasoning tasks, LLMs, execution modes, and execution policies, we find that multi-agent systems organized in a company-style hierarchy generally outperform other organizational structures. Besides, hierarchical coordination also reduces token consumption relative to flat collaboration in most settings. For example, for GPT-OSS-120B, the hierarchical setting improves performance over flat multi-agent system by 102.73% while reducing token usage by 74.52% on SQuAD 2.0. Further analysis shows that hierarchy helps most when tasks benefit from stable skill assignment, controlled information flow, and layered verification. Overall, our findings highlight organizational structure as an important factor in multi-agent reasoning, shaping not only effectiveness and cost, but also coordination behavior.

Abstract

Multi-Agent Systems

Explore similar work

OrgAgent: Organize Your Multi-Agent System like a Company

Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams

Organizational Principles Enable Collective Intelligence in Embodied AI