The Self-Driving Portfolio: Promise, Pitfalls, and the Practitioner Gap

• Institutional investors are exploring self-driving portfolios, where networks of AI agents continuously make and govern investment decisions.
• Research on real-world multi-agent systems highlights major risks, including error cascades, groupthink, self-modification failures, and weak auditability.
• Despite the promise of agentic investing, human oversight remains critical because AI systems are least reliable when markets are most stressed.

Slow deliberation is often the price of sound governance from the perspective of institutional finance. By the time investment committees meet and approve new decisions, markets and macro regimes have already moved on, and the opportunity behind the investment may already have withered or completely passed.

This finite bandwidth of human decision-makers is the central constraint in institutional asset management, Ang, Azimbayev and Kim argue in their paper, The Self-Driving Portfolio: Agentic Architecture for Institutional Asset Management (2026). Instead, they envision an extensive AI multi-agent system with specialised agents operating continuously critiquing one another’s work, adapting in real time, and rewriting their own code when errors emerge.

On one hand, the paper provides and ambitious view on the future of the financial industry from credible authors. On the other hand, a parallel body of literature on how multi-agent systems fail in production suggests the gap between this vision and a deployable institutional system is wider than it appears. This article examines both sides.

Both academic research and industry practitioners have explored a wide range of configurations for how agents might interact in portfolio management, producing an imaginative and rapidly growing body of work. What has been missing, until recently, is a unifying institutional framework that brings these ideas together in a form recognizable to a professional asset manager.

Research applying multi-agent reinforcement learning to portfolio management has grown rapidly. New architectures were fuelled by the premise that multiple agents working together is better than a single model. For example, MAPS (Multi-Agent Portfolio Management System; Lee et al., 2020) uses multiple independent agents acting like portfolio managers, each with its own investment style, and aggregates their outputs into a single diversified portfolio — effectively a portfolio of portfolios. Furthermore, MAMA (Multi-Asset Multi-Agent reinforcement learning; Kim & Lee, 2025) has a slightly different hierarchy where separate agents are used for different asset classes; some agents operate at the level of individual securities and others managing cross-asset relationships. Finally, He et al. (2025) introduce an additional layer of coordination with distinct agents responsible for macro conditions, firm-level fundamentals, portfolio construction, and risk control respectively.

Ang et al. (2026) add an institutional governance framework. They describe an organisation of approximately fifty specialised agents that are responsible for capital market assumptions and portfolios construction. These agents can critique and vote on each other’s outputs thus mirroring a formalized investment committee that operates continuously. For instance, a researcher agent proposes portfolio construction methods not yet represented in the system, while a meta-agent compares past forecasts against realized returns and rewrites agent code and instructions to improve future performance. The entire pipeline is governed by the investment policy statement that attempts to apply fiduciary and risk boundaries on the agent system.

A more skeptical body of research on how multi-agent large language model (LLM) systems behave in production suggests that the architecture described above may have some significant vulnerabilities.

A significant deterioration on the quality of the system could arise from the cascade problem when a single agent's mistake is amplified because downstream agents treat flawed outputs as truth (Reid et al., 2025; Zhu et al., 2025). For example, a an incorrectly interpreted policy statement that feeds into many outputs could corrupt the assumptions that every subsequent agent optimizes against. By the time this error reaches agent-based investment committee voting, all agents might think of this error as the real environment. This pattern is particularly common in long workflows with many sequential operations (Eswaran et al., 2026), which are precisely the conditions a fifty-agent investment pipeline would produce.

Ang et al. (2026) assume a system where agents critique and vote on each other’s outputs. Some studies, however, find that LLM agents acting in a reviewer capacity exhibit affirmation bias. Moreover, they tend to endorse outputs from producing agents rather than challenge them, particularly where agents share the same underlying model (Tian & Zhang, 2024; Roig, 2025). A pipeline of fifty agents reasoning from the same foundation model may therefore produce the appearance of independent scrutiny while converging on the first plausible answer. The evidence further suggests that adding reviewer agents does not reliably address this: Consensus tends to form more quickly rather than more accurately (Reid et al., 2025).

Ang et al. (2026) propose a meta-agent capable of rewriting its own code and prompts in response to past forecast errors. Yet the failure literature identifies “prompt drift” — situations in which seemingly minor modifications to agent instructions produce disproportionately large changes in system behaviour — as one of the most persistent challenges in deployed LLM systems (Vinay, 2025; Cemri et al., 2025). The risk may be especially severe during periods of market stress, when the pressure to adapt is greatest while the meta-agent must process a larger volume of noisier, less internally consistent information than during calmer market regimes.

Auditing decision-making by AI agents is genuinely difficult and requires a framework and additional oversight system (Reid et al., 2025; Raza et al., 2025). For example, European regulators request that investment decisions taken with AI systems are explainable and attributable to a responsible party (EU AI Act). Therefore, a system in which fifty agents critique and vote, and a meta-agent continuously revises their instructions, may face difficulty satisfying that requirement.

Finally, the portfolio literature itself identifies a limitation that the Ang et al. architecture does not fully resolve. LLM-based portfolio construction has been shown to perform well in stable market conditions but to deteriorate in high-volatility environments, a pattern that persists even in out-of-sample testing (Voronina et al., 2025). The self-improvement loop at the core of the Self-Driving Portfolio optimizes on past forecast errors, but market crises are by definition drawn from a different distribution than the historical record on which the system trains. A pipeline that is well-calibrated to the markets it has observed may therefore prove least reliable in the conditions that matter most to institutional clients.

The evidence for multi-agent and LLM-augmented portfolio construction is promising. The failure literature does not invalidate this, but it does suggest that the gap between a research prototype and a production-grade institutional system is larger than the paper acknowledges. The human overseer envisioned by Ang et al. cannot yet take a purely passive safeguard role.

Ang, A., Azimbayev, N., & Kim, A. (2026). The Self-Driving Portfolio: Agentic Architecture for Institutional Asset Management. arXiv, abs/2604.02279.

Bisconti, P., Galisai, M., Pierucci, F., Bracale, M., & Prandi, M. (2025). Beyond Single-Agent Safety: A Taxonomy of Risks in LLM-to-LLM Interactions. ArXiv, abs/2512.02682.

Cemri, M., et al. (2025). Why Do Multi-Agent LLM Systems Fail? ArXiv, abs/2503.13657.

Darwish, M., Hassanien, E., & Eissa, A. (2025). Stock Market Forecasting: From Traditional Predictive Models to Large Language Models. Computational Economics.

Eswaran, A., et al. (2026). CryptoAnalystBench: Failures in Multi-Tool Long-Form LLM Analysis. ArXiv, abs/2602.11304.

Hajaghaie, A., & Thulasiram, R. (2025). Leveraging Large Language Models and Retrieval-Augmented Generation for Enhanced Multi-Asset Portfolio Construction. IEEE CiFer 2025.

He, C., et al. (2025). Hierarchical AI Multi-Agent Fundamental Investing: Evidence from China’s A-Share Market. ArXiv, abs/2510.21147.

Heng, R., et al. (2025). Leveraging LLMs for Top-Down Sector Allocation in Automated Trading. ArXiv, abs/2503.09647.

Kim, S., & Lee, K. (2025). Multi-Asset Multi-Agent Reinforcement Learning for Portfolio Management. IEEE Access, 13.

Lee, J., Kim, R., Yi, S., & Kang, J. (2020). MAPS: Multi-Agent Reinforcement Learning-based Portfolio Management System. IJCAI 2020.

Raza, S., et al. (2025). TRiSM for Agentic AI. ArXiv, abs/2506.04133.

Reid, A., O’Callaghan, S., Carroll, L., & Caetano, T. (2025). Risk Analysis Techniques for Governed LLM-based Multi-Agent Systems. ArXiv, abs/2508.05687.

Roig, J. (2025). How Do LLMs Fail In Agentic Scenarios? ArXiv, abs/2512.07497.

Tian, C., & Zhang, Y. (2024). Collaboration Dynamics and Reliability Challenges of Multi-Agent LLM Systems.

Tong, X., Wei, L., & Yan, Z. (2026). Reliability Challenges of LLM Agents. Science and Technology of Engineering, Chemistry and Environmental Protection.

Vinay, V. (2025). Failure Modes in LLM Systems. ArXiv, abs/2511.19933.

Voronina, A., et al. (2025). Generative AI-enhanced Sector-based Investment Portfolio Construction. ArXiv, abs/2512.24526.

Vuković, D., Dekpo-Adza, S., & Matović, S. (2025). AI integration in financial services: a systematic review. Humanities and Social Sciences Communications, 12.

Wang, B. (2025). Empirical Evaluation of Large Language Models for Asset-Return Prediction.

Winder, P., Hildebrand, C., & Hartmann, J. (2025). Biased echoes: Large language models reinforce investment biases. PLOS One, 20.

Zhu, K., et al. (2025). Where LLM Agents Fail and How They can Learn From Failures. ArXiv, abs/2509.25370.