Mar 1, 2026 · Research · 9 min

The Fund That Runs Itself

In February 2025, a Sydney hedge fund called Minotaur Capital reported 13.7% returns over its first six months, beating 89% of global equity hedge funds tracked by Preqin. The fund has no analysts on staff. Its entire research operation runs on an AI system called Taurient that scans 35,000 articles a week, produces investment theses in minutes instead of days, and feeds positions to a three-person team. Across the Pacific, a 23-year-old former OpenAI researcher named Leopold Aschenbrenner ran Situational Awareness LP to a 47% return in the first half of 2025, six times the S&P 500. And in San Francisco, Numerai closed a $30 million Series C at a $500 million valuation, with J.P. Morgan committing $500 million in capacity to a fund whose trading signals come entirely from a tournament of machine learning models built by anonymous data scientists around the world.

Something fundamental is shifting in how capital gets allocated. The question is no longer whether AI will reshape asset management. It's whether the fund of the future has a portfolio manager at all.

The performance gap

The data is getting uncomfortable for traditional firms. AI-first funds aren't keeping pace with the market. They're pulling ahead of it.

The broader numbers confirm the trend. Over 70% of global hedge funds now use machine learning somewhere in their trading pipeline. More than 35% of new fund launches in 2025 branded themselves as AI-driven or AI-enhanced. Advanced AI strategies outperformed traditional quant funds by 4-7% in 2024, and firms incorporating generative AI into decision-making clocked 3-5% better returns across equity strategies where pattern recognition matters most.

The industry recorded its best year since the 2009 financial crisis recovery, with 12.6% average returns across all strategies. The returns were driven by stock-picking and macro strategies where AI-driven pattern recognition delivers the most alpha. The performance gap between AI-native funds and traditional managers is widening, and capital is following. The hedge fund industry enters 2026 with its highest inflows in almost two decades, based on a survey of 340+ investors representing $7.8 trillion in AUM.

Anatomy of an autonomous fund

What makes these funds different isn't that they use AI. It's how they're structured. A traditional fund employs 10-50 analysts who read filings, build models, attend management meetings, and produce investment memos over days or weeks. An AI-first fund replaces most of that pipeline with agent systems that operate continuously.

Minotaur's architecture is the clearest example. Taurient operates as a multi-agent system. One set of agents ingests and scans roughly 35,000 news articles per week across global markets. Another layer interrogates those articles, extracting signal about listed companies and flagging opportunities that have slid under the market's radar. A third layer produces rapid investment assessments, compressing a five-day human research process into minutes. Three people operate the entire firm. The AI costs about half the salary of a single junior analyst while replacing 20 of them.

graph TB
  DATA["Market Feeds · News · Alt Data"] --> R["Research Agents"]
  R --> A["Analysis Agents"]
  A --> K["Risk Agents"]
  K --> P["Portfolio Construction"]
  P --> T["Trade Execution"]

  style R fill:#f5f5f4,color:#0a0a0a,stroke:#0a0a0a
  style A fill:#f5f5f4,color:#0a0a0a,stroke:#0a0a0a
  style K fill:#f5f5f4,color:#0a0a0a,stroke:#0a0a0a
          

This architecture maps directly onto the protocol stack that is standardizing across the industry. MCP provides tool connectivity: agents accessing market data feeds, financial databases, SEC filings, and broker execution systems through standardized interfaces instead of custom integrations. A2A enables multi-agent coordination: research agents, risk agents, and execution agents communicating through structured task lifecycles with built-in auditability. The agent harness orchestrates the full pipeline, enforcing guardrails, managing handoffs between specialists, and maintaining the compliance trail that regulators require.

Numerai takes a radically different approach but arrives at the same destination. Instead of building proprietary AI research, Numerai crowdsources predictions from thousands of data scientists worldwide. Participants stake NMR tokens on their models, creating a financial incentive that selects for accuracy and penalizes overconfidence. The fund aggregates predictions into a stake-weighted meta model that drives trading. Over 2,000 LLMs now interact with Numerai's APIs autonomously. Assets grew from $60 million to $550 million in three years, and J.P. Morgan's $500 million capacity commitment signals institutional validation of the model.

The incumbents respond

This isn't happening at the margins. Every major financial institution is building agent infrastructure into its investment operations.

BlackRock's Aladdin platform, which generates over $4.4 billion in annual technology services revenue, launched Aladdin Copilot with generative AI capabilities embedded across its investment management stack. In October 2025, Aladdin Wealth introduced Auto Commentary, an AI tool that converts complex portfolio analytics into advisor-ready insights. Morgan Stanley was the first to implement it across its Portfolio Risk Platform. BlackRock's acquisition of Preqin brings private markets data, notoriously unstructured and fragmented, into the Aladdin AI ecosystem.

Wells Fargo partnered with Google Cloud on Agentspace for AI-powered FX post-trade processing. JPMorgan, running a $17 billion technology budget, expanded from 450 AI use cases in 2023 to over 1,000 by 2025, deploying agents across trading, risk analytics, compliance, and client services. Citadel debuted an AI research assistant for its equities division in December 2025.

The startup ecosystem is moving just as fast. Public's Agentic Brokerage lets retail investors create custom stock indexes through text prompts, with over 2,500 AI-generated strategies created in the first month. A full AI-powered portfolio manager is planned for 2026. SoFi launched the Agentic AI ETF (AGIQ), tracking companies building autonomous AI infrastructure. Minotaur Capital is preparing its own ETF to bring AI-driven equities management to retail investors.

The regulatory reckoning

The SEC's 2026 examination priorities, released November 2025, signal that regulators understand what's happening. Emerging technology and AI are explicitly listed as focus areas. Examiners will evaluate automated investment tools, algorithmic models, and AI-based systems for accuracy of representations and alignment with regulatory expectations. The core question is pointed: do AI-driven recommendations meet fiduciary standards?

Under Section 206 of the Investment Advisers Act, the duty of care and loyalty does not change because AI is in the loop. An adviser who deploys an autonomous agent to make portfolio decisions still bears full fiduciary responsibility for those decisions. But the opacity of agent systems creates enforcement challenges that existing frameworks weren't designed for. A recent SEC enforcement action indicated that failure to ensure reliability of automated trading models or implement written policies governing them could constitute a breach of fiduciary duty.

FINRA's 2026 oversight report addresses autonomous AI agents directly, flagging risks around unchecked autonomy and agents acting beyond a user's intended scope. The report recommends supervisory processes specific to each agent type, including human-in-the-loop protocols and behavioral guardrails. Securities class actions targeting AI misrepresentations doubled between 2023 and 2024. In 2025, the DOJ joined the SEC in pursuing "AI-washing" cases with parallel civil and criminal actions.

The uncomfortable gap: 57% of financial services organizations are still developing the internal capabilities to fully leverage agentic AI. Only 9% of investment firms reported using agentic AI in Q4 2025. But 44% of finance teams plan to deploy agents in 2026, a 600% increase, with private equity firms at 95% planning adoption. The regulatory framework is chasing adoption that moves faster than any rule-making process can respond.

The trajectory

Four forces are converging, and they compound.

First, performance creates a self-reinforcing cycle. AI-first funds outperform. Capital flows to outperformers. More capital enables more data, better models, and faster iteration. Traditional firms either adopt or lose allocations. As more funds adopt, the edge accrues to those with superior data, better implementation, and stronger risk management.

Second, the cost structure is collapsing. Minotaur runs with three people what used to require 20 analysts. The AI costs roughly half a junior analyst's salary. Wealth advisory firms deploying agents report 40-50% reductions in portfolio operations costs and 30-40% AUM growth per advisor. As inference costs continue to drop, the economic advantage of AI-first operations widens.

Third, the protocol infrastructure is maturing. MCP, A2A, and agent harnesses provide standardized building blocks for autonomous fund operations. A fund no longer needs custom integrations with every data provider, broker, and compliance system. It connects through protocols. The same architecture that standardized how agents access tools and coordinate with each other now standardizes how capital gets managed.

Fourth, retail access is opening. Public's Agentic Brokerage, Minotaur's planned ETF, and Numerai's institutional partnerships all signal that autonomous investment management is moving beyond the institutional world. Autonomous portfolio management at institutional quality, delivered to individual investors, at a fraction of the cost.

The question facing every asset manager, allocator, and investor is not whether autonomous funds will work. The performance data already answers that. The question is how much of the value chain gets automated before the remaining human role becomes purely supervisory: setting objectives, defining constraints, and reviewing outcomes while agents handle everything from research to execution to compliance. That transition is not coming. It's underway. And the returns speak for themselves.