In February 2025, I published an essay in Institutional Investor titled "Investment Management in a Box." The argument was straightforward, if uncomfortable: artificial intelligence had reached a point where you could, in theory, replace nearly every function of an asset management firm — research, portfolio construction, risk management, compliance, even the CEO — with specialized, autonomous, personalized AI agents that communicate, learn from one another, and improve over time.
The evidence I cited was a thought-provoking paper in which researchers built a simulacrum of a hospital called Agent Hospital, where the patients, nurses, and doctors are autonomous LLM-powered agents. The agentic physicians achieved 93 percent diagnostic accuracy after treating 10,000 simulated patients, outperforming their human counterparts in the same dataset by six percentage points.
I concluded the essay with what I intended as a rhetorical challenge: the asset management industry's stubborn preference for keeping AI as a co-pilot rather than a pilot was a choice, not a necessity.
Then, recently, I encountered something that required me to revise my thinking entirely.
Someone had already built the “Box.” His name is Dr. Miró Mitev, and he is the founder and CEO of the Swiss asset manager Smart Wealth Asset Management AG. (I was recently named an advisor to Smart Wealth but receive no compensation for this role and hold no financial interest in the firm or the fund. This proximity has given me the access needed to write this essay in detail.)
Hiding in Plain Sight
I track managers deploying transformative technology in institutional investing. Lately, I’ve come across many managers claiming to use AI agents. Scratch the surface, and you’ll find the majority are using agents to improve operational efficiency, not make autonomous investment decisions.
So, when a colleague forwarded me a description of Smart Wealth's autonomous, multi-agentic investment process, my first reaction was disbelief. What Smart Wealth had built and deployed was not a theoretical construct; it was the system I had argued, in print, that someone should build.
Someone already had. My second reaction followed immediately: "And they’ve been running it since when?"
The technology underlying Smart Wealth’s multi-agentic systems has its roots in Mitev's 2000 master’s thesis and his 2002 doctoral dissertation on applying neural networks to capital markets at a moment when most of the industry had not yet heard of machine learning, let alone large language models. From 2001 to 2009, at a major European industrial conglomerate, he led a multidisciplinary team of electrical engineers, computer scientists, physicists, mathematicians, and data researchers, none of whom had backgrounds in finance, in building one of the world's first fully automated AI-driven systems for forecasting and optimizing energy supply and demand.
The success of that system led Mitev to propose something bolder to the company: use an iteration of the same system to manage the conglomerate's pension fund, including forecasting asset prices and optimizing the portfolio over time to meet pension liabilities within a regulatory framework. The suggestion was accepted, putting the technology to work on an actual investment portfolio.
What Mitev and his team had built, though the term was not yet in common use, was a system of autonomous AI agents capable of making forecasting and optimization decisions independently within a collaborative framework, without human intervention at the point of the decision.
At the time, in the early 2000s, much of the industry still treated investment edge as something derived from human judgment, relationships, and qualitative thesis-building.
Mitev and his team rejected this premise entirely. In their view, capital markets are, at their core, large-scale optimization and pattern-recognition problems operating under uncertainty. They believed the same advanced AI systems being developed for voice recognition, image classification, and industrial process control could be applied to markets too. And the way to solve these problems was to start, not with human constructs, but with data and let the machines find the relevant patterns and relationships. For a team of electrical engineers, physicists, and computer scientists, this was not a radical insight. It was the obvious one.
A Reuters story on how the conglomerate benefited from the technology Mitev and his team built caught the attention of institutional investors around the world. Sensing a burgeoning commercial opportunity, the company began licensing the IP to institutions including JPMorgan, Société Générale, Banco Santander, and Nomura, all of which used it to generate custom signal forecasts and optimize their portfolios.
Mitev and his team continued to develop the system, and in 2016, he decided to found Smart Wealth in Zurich and use the system exclusively to manage institutional assets.
Strip away the chronology and what you have is an autonomous AI agentic system developed two decades ago that has evolved from forecasting and portfolio optimization into Smart Wealth’s current autonomous system. Not a demo. Not a research paper. Not a back test. And not built on large language models.
Building Agents without LLMs
How did Mitev build autonomous agents at least a decade before "Attention is All You Need" gave rise to transformers and the generative AI era? The answer begins with what his system is not. Large language models, the architecture underlying ChatGPT, Claude, and most of what the public now associates with AI, are built to process and generate language. They excel at synthesizing information, drafting text, and simulating reasoning in human-readable form. That makes them powerful research assistants. But Mitev argues they were not designed to solve the core problem of portfolio management, which is not a language problem but a decision problem: how to allocate capital under uncertainty, across time, in a non-stationary environment where the rules change continuously.
Mitev's system was designed from the ground up to solve that problem directly. It is a multi-agentic system, but architecturally distinct from LLM-based approaches. It does not use transformers to forecast prices, does not route decisions through a chain of GPT instances simulating portfolio managers, and does not depend on any third-party foundation model or external vendor API. In fact, the model relied on machine learning and statistical techniques that long predated LLMs.
Instead, the forecasting architecture combines discriminative, supervised, and unsupervised machine learning models — feedforward neural networks, dynamic factor models, vector autoregressive models with exogenous variables, and multi-regression models — trained on numerical financial time series data using techniques whose performance can be tested, validated, and audited. Diversified by design: each model class has distinct failure modes, and the combination reduces reliance on any single component.
Feature selection is sparse. Unlike traditional quant models, which can rely on a large library of hand-crafted signals and factors chosen by human analysts based on prior research, intuition, or market experience, Mitev's system is deliberately lean in its feature engineering. The system is designed to discover structure in the data rather than have humans impose it.
The system uses genetic algorithms to search roughly 2,000 publicly available macroeconomic, fundamental, and technical indicators and identify, by asset, a small subset that the system determines has predictive value. He then uses a second genetic algorithm to perform thousands of iterations per cycle on this proprietary forecasting data to find optimal portfolio weights for a given risk target. The system then decides autonomously when to rotate, when to hold cash, and how to rebalance against transaction costs. No investment committee. No overrides.
These models do not hallucinate because they do not generate; they classify, regress, and search. They do not collapse under stress because the supervised learning algorithms and fitness function anchor outputs to labeled target data, not output generated by an LLM. And they do not exhibit the kind of agentic misalignment that concerns critics of LLM-based systems, because each agent operates within a narrow, mathematically defined objective and explicit constraints, not a free-form natural-language goal that drifts under pressure.
Over the past twenty years, this core architecture has been refined, retrained, scaled, and optimized many times, and yet, even with the advent of LLMs, it remains fundamentally unchanged. What has changed is the breadth of inputs, the speed of execution, the scale of the investable universe, and the granularity of risk control. That last dimension marks a significant evolution: in the system's early years, forecasts and portfolio recommendations were delivered to clients, but execution and risk management remained outside the technology. At Smart Wealth, the system was extended to automatically execute investment decisions and dynamically manage portfolios on a daily basis, embedding risk control directly into the process rather than leaving it to a separate human-managed step.
How Does This Differ From Quant Investing?
Traditional quant firms typically develop and deploy multiple independent models to perform different tasks (e.g., alpha generation, risk estimation, portfolio construction, execution optimization), with outputs feeding into one another sequentially, in parallel, or hierarchically. Some use ensemble approaches, combining signals by averaging, weighting by recent performance, or using a meta-model to allocate across them. Others use regime-switching models or multi-factor models that stack signals such as value, momentum, quality, and low volatility into a combined score.
The critical distinction between traditional quant models and Mitev's system is not architectural complexity; it is the locus of decision-making. Quant models, however sophisticated, are originally conceived and circumscribed by human intelligence. They are deterministic, not self-learning, and their capacity is inherently limited by the assumptions built into them at design. They do not make forecasting and optimization decisions independently. A human remains in the loop at the point of decision.
Mitev's system does not. His agents work independently and collectively, forecasting, optimizing, and executing without human intervention at the moment of decision. That is not a refinement of the quant model. It is a different category of system.
Six Agents Running the Portfolio
In my December essay, I imagined sixteen agents running an entire asset management firm. Mitev does it with six mission-critical core investment-related agents.
- The Scout screens roughly 5,000 globally listed securities, using Principal Component Analysis and Cluster Analysis to identify genuine diversifiers and discard redundancies.
- The Detective searches a pool of 2,000 macroeconomic, fundamental, and technical indicators, using genetic algorithms to identify the factors with predictive power and quietly discard the rest.
- The Forecaster runs a dedicated ensemble forecasting engine for every asset using neural networks, dynamic factor models, vector autoregressive models, and other techniques, with every forecast back-tested, validated out-of-sample, and live-tested before it influences portfolio weights.
- The Strategist runs thousands of iterations of a genetic optimization algorithm each monthly cycle, seeking the highest risk-adjusted return at a given volatility target, and rotating into cash without a committee meeting when forecasts warrant it.
- The Trader routes the Strategist's target weights through a proprietary execution and reconciliation system, with compliance checked along the way and trades reconciled automatically.
- The Watchdog runs continuous compliance checks, produces daily automated factsheets, and maintains an audit trail.
Six agents. No redundancy. No override of individual decisions.
Critically, the power of this system rests in the collaboration between agents. Each agent publishes its output to a shared internal data layer. The Scout's universe feeds the Detective. The Detective's indicators feed the Forecaster. The Forecaster's signals feed the Strategist. The Strategist's weights feed the Trader. The Trader's executions feed the Watchdog. The Watchdog's risk signals feed back into the Strategist as constraints for the next cycle. A closed, continuous loop with zero human interference.
Humans, however, haven’t disappeared. They supervise the system, rather than make individual investments decisions. Embedded within every agentic workflow are FC-workers, lightweight code components that function as dispatchers and monitors, verifying that each agent completes its assigned tasks on time, without errors, and without bottlenecks. Should a bottleneck occur, the responsible FC-worker routes pending tasks to one or more agents for immediate resolution. Any workflow failure, such as missing or delayed data, triggers an immediate alert to a human supervisor while the agent attempts to reload the data and continue the calculation. If the problem persists, the FC-worker escalates it to a human officer who investigates and resolves it. The loop runs without human intervention under normal conditions; the moment it cannot, a human is in the chain.
Smart Wealth's investment process is designed to operate autonomously. Fiduciary responsibility, however, is not delegated to the machine. As a regulated asset manager in Switzerland, Smart Wealth bears full legal and fiduciary responsibility for all investment outcomes and remains entirely accountable for their impact on client portfolios.
What Is Fundamentally Human About Investing?
I began this essay arguing that someone could, in principle, build the Box. I was wrong about one word. Not could. Already had.
Mitev did not set out to disrupt asset management. He set out to solve an engineering problem, applied the tools his training gave him, and assembled a team of engineers and scientists who approached investment problems from outside the financial tradition. They believed the tools they used every day in other fields could also work in markets and followed the logic where it led. What he built is, to my knowledge, one of the only examples of autonomous agentic investment management currently in operation that has been accountable to regulators and auditable at every step for more than two decades.
I closed my December 2024 essay with a line I borrowed from Michael Taylor of Coldwater Economics: "My starting point is that, one way or another, investing is and will remain a fundamentally human activity." I disagreed with this sentiment then. Eighteen months later, I disagree with it even more.
The question is no longer whether autonomous, multi-agent investment management is possible. One answer to that question has been running quietly and continuously in Vienna beginning in 2000 and then in Zurich since 2016. The question now is whether the rest of the industry is willing to recognize what has already been demonstrated, and to think seriously about what it means.
