The last time markets changed at the root, machines took over the trading — execution, market-making, and the systematic strategies that decide what to buy and sell. What stayed human was the reasoning behind big decisions. This time that goes to the machines too — not just the trade, but the research, the risk, the allocation.
The algorithmic takeover ran on speed. Regulation NMS, in force by 2007, split trading across dozens of competing electronic venues, each bound by its trade-through rule to honor the best price posted anywhere — so being first to that price was the whole game. High-frequency trading climbed from about 20% of U.S. equity volume in 2005 to about half by 2008 and roughly two-thirds by 2009, and has stayed a large share of volume since.1 The machines did the job well: faster execution, tighter markets, and systematic strategies at scale — deciding where and what to trade. What none of it could do was reason — given new, unstructured context, decide on its own how to trade over long horizons. Call that machine reasoning; the last era ran without it. The early 2000s gave us statistical models very capable of finding patterns in data and trading on them — not of forming a view on what a move meant, gathering more data to test it, and acting on that view. Trained on much the same data, funds ended up with much the same models — picking up the same signals and the same noise. When one crowded book unwinds, the rest go with it — the August 2007 Quant Quake, quant funds built on the same factors, all deleveraging together with no news to explain it. Call it the algorithmic herd.
A reasoning agent is a categorically different machine.2 It does not detect faster; it holds a view, weighing what a move means based on new information before it acts. Done right, it is the first machine that can decline the herd instead of joining it. Done wrong, it builds a new herd of its own.
I've built trading systems in both paradigms, and today I run agentic ones against live markets. Everything in this essay is written from that seat.
The last era wasn't wrong — it used the best technology it had, and took it as far as it would go. The only wrong move now is refusing the technology that replaced it. To see why this era is different, start with the limits of the last one. Machine learning for trading used to mean a stack of supervised models, heavy on feature engineering, trained on time-series data for pattern recognition: a sell signal, a regime change, a mispriced option. Some unsupervised models here, some optimal-control algorithms there — mostly independent pieces wired together, each with a specific function. A classifier has no concept of what a headline means. It scores a position; it does not reason about one. It flags an anomaly; it does not reason about its cause. It gives one answer; it does not loop back for better tools or data. Wherever finance reached for machine learning, it got a scorer, not a reader. Call it the detection era.
The old machines could score the world. They could not hold a view of it.
Herding is not the only way the old machines break. In 2025 one of the most storied quant funds alive had its worst year in over a decade — flagship strategies down double digits in a single month — because its models kept pressing bets a market remade by the AI trade simply ran over.3 The patterns were real. The regime was new. Detection has no answer for a market it has never seen.
What changed is that machine reasoning finally arrived in general-purpose form — large language models that reason by spending inference-time compute to search over chains of thought, and agents built on them that observe, think, act, and react in a loop.4 AI researchers had pursued this for decades — first through symbolic AI, later through reinforcement learning — without a result that generalized. Whether the frontier reasoning models five years out are still LLMs underneath is an open question. There is no doubt they will be more capable; whether they are also safer is a question of governance, not time. In 2026 we got a different kind of machine, one that quietly breaks the frame the quants spent two decades building.
The shift is reasoning, not speed
The old race was latency and detection: be first to the price, first to the fill, first to flag the pattern everyone else would trade a beat later. At those horizons the winner did not need a view — only a reflex. That race was already ending before agents arrived: deep learning had been pulling strategies from high-frequency toward mid-frequency horizons, where prediction quality matters more than latency. Machine reasoning carries the migration the rest of the way, to horizons where the contest is interpretation — an agent that weighs a central-bank meeting against a supply chain against a buried analyst note, verifies its read and digs deeper on its own, and carries the position on that argument, with its own dynamic stop-loss and its risk sized to it. Whatever the last era's models could do, they never went and gathered the data they weren't handed — an agent does that unprompted, and acts on what it finds. Speed decided who saw first. Machine reasoning decides who understands first.
This does not stop at quant. Fundamental investing — the analyst reading an earnings call, mapping tone to a view — was long treated as the last redoubt of human reasoning, and an agent can now do a meaningful share of it under human verification. For the first time, quant and fundamental sit on the same substrate. That division was never epistemic but technological — different questions needed different machines, so firms built different desks. One machine that reads, reasons, and acts erases the reason to keep them apart.
The stack just collapsed
The same shift has another face: effort. It didn't get destroyed — it transformed. Standing up a trading system used to mean a deep pipeline of specialized models wired together: a team, a year, a quant PhD or three. A frontier LLM collapses much of that into a single reasoning loop directing retrieval, tools, evals, and verifiers. What took a desk a year, a builder now prototypes in less than a week — a raw system, albeit probably not a profitable one and definitely not yet a safe one.
The cost did not vanish; it relocated. The week gets you the prototype; the year moved into inference compute at scale and the evals, monitoring, limits, kill switches, and audit trails — i.e., the agentic governance — that stand between a demo and a system that can hold capital. The specialist mix shifts with it. What collapsed is the cost of the first mile, not the last. One consequence follows directly: when capable agents get cheaper to build, they proliferate, and the market fills with new artificial participants — a view that once took a research team and a mandate can now be formed and acted on by a system built in a week.
Agentic proliferation cuts both ways. Detection strategies crowded because they trained on much the same data. Agents crowd for the same reason: built on the same short list of foundation models and pointed at the same feeds, they can reach the same view at the same moment. The first herd chased trends on shared momentum signals; the agentic herd can chase them on shared weights.
A market of reasoning agents is still a herd if every agent reasons from the same weights.
What is striking about generative AI is how many of the detection era's problems it surfaces and amplifies — crowding, correlated failure, systems that move in lockstep — and that none of them are unique to finance. AI safety is the older discipline here: researchers worked on model monoculture and correlated failure for years before any of it reached a trading desk, and beyond finance the stakes run higher still. Proliferation did not create the risk so much as make it urgent — more agents reasoning from the same handful of models means the old failure mode arrives faster and wider. Regulators are already simulating that failure mode. At the ECB's Sintra forum in June 2026, Bank of England Deputy Governor Sarah Breeden warned that autonomous AI agents could amplify volatility under stress and, in the extreme, drive a market meltdown, and that existing regulation was never built for agentic AI; the Bank's Financial Policy Committee has flagged firms' concentrated dependence on a few AI providers as the channel through which exactly this same-weights herd forms.5
A reasoning agent can decline the old detection herd — yet pointed at the same foundation models as everyone else, it only joins a new one. You change herds; you do not escape herding. That is the clue to what the industry gets wrong: the case for trading agents runs by analogy to software — agents learned to write code, so they will learn to trade, and whoever ships first is handed an edge.
A trading agent is not a coding agent
Consider what a coding agent actually does. It works in a sandbox against a fixed target — your goal stated in a prompt, your deterministic codebase, and a deterministic test suite — writing, running the tests, reading the failures, and rewriting until they pass. A non-deterministic agent handles this well because every attempt is scored against an unambiguous grader: the tests pass or they do not. Whatever the engineering behind it, the problem it is solving is a Markov decision process (MDP). A trading agent has none of that. It acts in a non-deterministic market with no fixed oracle to grade it, never observing enough of the market to pin down its state, let alone predict it. It is solving not an MDP but a partially observable MDP (POMDP).6
Then, of course, add all the other market participants, themselves adapting and reacting — some now reasoning, some oblivious, most still just fast detection models. No test stays passed. The agent's own trade moves the price, and the moved price becomes the next agent's input — just as with human participants. Run that across a market full of agents, all reacting to one another, and the feedback can explode. The problem changes class once more: a multi-agent, partially observable stochastic game — game theory, now played by machines. In the words of Pedro Domingos: figure out how a large multi-agent system can autonomously coordinate to maximum effect, and you'll win a Nobel Prize, a Turing Award, and a trillion-dollar fortune all in one.7
The good news is that the edge does not wait on an optimal solution — quant finance has never had one. We do not need to solve the multi-agent game to find an edge, or a better one than we have now. What we do need is to recognize that a trading agent is not a coding agent, and that the difference is structural, not cosmetic — it demands a markedly different approach to deployment, risk, simulation, and above all backtesting.
None of that means the off-the-shelf model wins. A frontier model, dropped unmodified into an agent loop, gets a market right no more reliably than a vanilla classifier did — reasoning is not a shortcut. The premise underneath it is far more powerful, though, and not one you can afford to ignore. Why a trading agent is not a coding agent — and what separates the firms that got it from the ones that didn't — is an essay of its own.
The era is wider than trading
Everything above is framed around the trade because the trade is the hardest version of the problem — a POMDP played inside a multi-agent game, where the target moves and everyone else is adapting to you. Little else in finance is that adversarial, and part of using this technology well is judging which problem you are actually in. The same collapse runs through the research floor, the risk book, the credit desk. The interpretive work that defined the analyst's craft — the initiation note that took weeks, the deep-dive on a name nobody else covered, the screen across a thousand companies — now compresses into a fraction of the time it took. The machine that reads one earnings call can read the whole sector and surface the name that breaks from consensus. The agentic era does not arrive desk by desk. It arrives as a substrate under all of them.
Where this goes next
The prize moves, and not to where the crowd is looking. Not to the signal — a signal gets crowded and arbitraged the moment it works. Not to the model everyone can rent, either — same weights, same shelf, same reasoning for anyone who pays. It moves to the best reasoning agent — the best context, the best adaptation, and the reliability and governance to hold capital: an agent that reaches for its own tools, fetches its own data, spins off its own sub-agents and specialized models, checks its own work, improves from its own experience, and acts within limits it can be trusted to hold. That advantage is not privileged access to something no one else can see — it is the data you own, the tools you built, the governance you run. We have come a long way toward that agent in a few short years, and we are not there yet. Much of the work that remains is not the foundation model's to do; it falls to the verticals — the firms building on top — to carry that last mile. The deeper shift is not in the signal at all. It is in how these systems are built and experimented on, and the firms that see it early are already building for it. I count myself among them. What that shift is — and how to build for it — is where this series is headed.
You don't fight a wave this size
Most quants I talk to love LLMs about as much as floor traders loved quants in the 1990s — and the charge is the one the pits once threw at them: that's not real trading, it doesn't understand the market. The drawbacks behind that charge are real, and I have lived them: same prompt, different answer; no error bars; a model that changes underneath you. Anyone who has put one of these systems near live capital has felt all three — I've hit every one. They are not a reason to stand back; they are the whole reason the hard work moved to evals and governance. Control and reliability are not properties you wait for in a model release; they are properties you build around it.
The older doubt I owned too. In 2023 the technology hallucinated and could not do arithmetic, and it was a consumer product — I believed a version of that once. Recently I sat down to test the frontier models the honest way. I took a quant paper fresh off arXiv, posted after the models were trained, its main theorem newly proven. Without showing the model the paper, I posed a problem it could only answer by re-deriving that theorem itself. It did; I checked the steps, and they held. Not recalled — reasoned. Something changed.8
This is not an argument to win — eras do not wait to be believed; they change hands. The last one passed from the floor traders to the statisticians who out-modeled them. This one passes to the AI-natives: the people who treat models as things you train, evaluate, and govern — not products you rent. When the wave is this size, you don't fight it. You surf it.
Train your own models
Not everyone is standing back — the brightest quants are already all-in. DeepSeek, the general-purpose lab that reached the reasoning frontier and now ships some of the best open-weight models in the world, was spun out of High-Flyer, a Chinese quant fund. That is the part I find most exciting: the last real step-change in open-weight AI came from a team forged in quant finance, and the next one may too — the field has the talent, the compute, the discipline, and every incentive.
Which brings me to a point I will keep coming back to. The target state — the agents that are actually yours — lives where you train and post-train your own models. I hope the frontier labs shift from owning the one model everyone runs toward building enterprise tools to train them more efficiently — closer to what OpenAI was before ChatGPT. That keeps your model and your data under your own control. It is also where the edge is — the one way out of a market where every agent reasons from the same weights.
This time the pressure will not wait for anyone to come around. It comes straight from the market. On the trading desk it shows first: the detection models that carried the last era begin to lose to the funds built for this. Everywhere else it wears a different face — the analyst who now covers ten times the names, the credit book that underwrites in a day at a fraction of the cost, the risk group that answers before yours has pulled the data. The mechanism differs by desk. The verdict does not.
I'll keep building the public thesis here in Perspectives — the frames, the theory, the why — in full and for free. Subscribe to get each signal as it publishes. AgenticAlpha is where the theory becomes operational: the specific AI models and infrastructure, the industry trends I'm tracking live, and where real advantage is forming and eroding across the AI landscape. If Perspectives is the map, AgenticAlpha is the terrain.
Subscribe to AgenticAlpha on Substack →
Welcome to Perspectives. This is the first signal. There will be many more.
Footnotes
-
Figures track high-frequency trading's share of U.S. equity volume: about 20% in 2005, roughly half by 2008, peaking near 60–70% around 2009–2010, then easing back toward half by mid-decade — year-by-year figures vary by source and definition (TABB Group, Celent, SEC). Automated order handling more broadly — every algo-routed order, not just the fastest — runs higher still and today touches the large majority of U.S. equity orders. The proximate trigger was Regulation NMS. Its Order Protection Rule — the "trade-through" rule (Rule 611, effective 2007) — barred any venue from executing through a better price displayed elsewhere; its Access Rule (Rule 610) capped access fees and seeded the maker-taker rebate model. Both made speed the battleground, though the deeper driver is the continuous limit-order-book design itself, which turns speed into mechanical arbitrage regardless of any single rule (Budish, Cramton & Shim, 2015). Decimalization (2001) — penny ticks in place of sixteenths — was the earlier precondition that collapsed spreads and left high-speed, high-volume market-making as the only profitable model. The two canonical machine-driven dislocations differ in mechanism but share the lesson: the August 2007 "Quant Quake" was crowded quant strategies deleveraging together — a decision-level herd (Khandani & Lo, 2011) — while the May 6, 2010 "Flash Crash" was a liquidity withdrawal at the execution layer. ↩
-
AI reasoning has two long research lineages. Symbolic AI — formal logic, automated theorem proving, planning, and expert systems (Newell & Simon's Logic Theorist, 1956; the expert-systems boom of the 1980s) — encoded reasoning as explicit rules over symbols and stalled on brittleness, the knowledge-acquisition bottleneck, and combinatorial explosion. Reinforcement learning produced powerful but narrow results in bounded settings (TD-Gammon, 1992; Atari, 2013; AlphaGo, 2016), sample-hungry and poor at generalizing beyond the task it trained on. The scalable general-purpose reasoner came from neither alone: a pretrained language model prompted to reason step by step (chain-of-thought; Wei et al., 2022), given inference-time compute to do it (OpenAI's o1, 2024), and post-trained with reinforcement learning against automatically verifiable rewards in math and code (DeepSeek-R1, 2025). Reinforcement learning was not the wrong tool so much as the wrong target — until it was pointed at verifiable rewards on top of a language-model prior. This reasoning is probabilistic in a specific sense: the model defines a distribution over next tokens, so a chain of thought is a sample from a distribution over reasoning traces, not a deterministic derivation — the same prompt can yield a different trace and a different answer. "Inference-time search" is drawing and scoring many such traces (temperature setting how sharply the distribution is peaked; self-consistency marginalizing over them, Wang et al., 2022). This is not the older subfield that "probabilistic reasoning" usually names — Bayesian networks and graphical models over explicit random variables (Pearl, 1988) — which is a different object. ↩
-
In 2025, the flagship public funds of a leading quantitative firm posted their worst annual results in over a decade — including double-digit losses in a single month — as factor and volatility models were caught offside by a liquidity- and sentiment-driven, AI-fueled rally. ↩
-
Today's agentic systems, trading agents included, are almost without exception built on large language models — a contingent fact, not a law. A parallel line of research builds agents on learned world models — systems that learn a predictive model of their environment and plan against it — which some prominent researchers argue is a more principled route to grounded reasoning and planning than next-token prediction (Ha & Schmidhuber, "World Models," NeurIPS 2018; LeCun, "A Path Towards Autonomous Machine Intelligence," 2022; Hafner et al., DreamerV3, Nature, 2025). In practice that program has produced strong results in bounded, simulated settings but has not yet matched LLM-based agents in general capability. LLM agents, by contrast, have shown real competence on open-ended work — most visibly coding agents, which now resolve a large and rising share of real GitHub issues on the human-validated SWE-bench Verified benchmark (Yang et al., SWE-agent, NeurIPS 2024; Jimenez et al., SWE-bench, ICLR 2024). For now, the frontier of practical agentic capability runs through LLMs; whether it stays there is open. ↩
-
Sarah Breeden, Deputy Governor for Financial Stability, Bank of England, at the ECB Forum on Central Banking, Sintra, 30 June 2026: autonomous AI trading agents could amplify volatility in periods of stress and, in the extreme, contribute to a market "meltdown," and existing regulation was not designed for agentic AI. The Bank's Financial Policy Committee has separately identified concentrated dependence on a small number of AI models and providers as a channel that can synchronize behavior across firms with different mandates. Bank of England survey work in 2026 also found a majority of surveyed firms already using AI, with autonomous trading agents moving into production. ↩
-
A Markov decision process (MDP) is the standard mathematical model of sequential decision-making under uncertainty, from dynamic programming and optimal control (Bellman, 1957): at each step the agent observes the full state, acts, and that state captures everything needed to decide what comes next. The model is independent of any solver — reinforcement learning is one family of methods for solving MDPs, and an agent faces an MDP whether or not it was trained with one. A partially observable MDP (POMDP; Åström, 1965) drops the full-observability assumption — the agent never sees the true state, only noisy, incomplete signals of it, and must act on a belief about where things stand; that is far closer to trading, where no participant observes the whole market (more precisely, incomplete information — private information and adverse selection; Kyle, 1985; Glosten–Milgrom, 1985). With many such agents adapting to one another, it becomes a multi-agent, partially observable stochastic game (its fully-observed form is the stochastic game of Shapley, 1953, reintroduced in reinforcement learning as the Markov game, Littman, 1994; here with full observability dropped) — harder still, and where the game theory enters. The jump is one of solution concept, not just difficulty: an MDP has an optimal policy; a game against others who are themselves optimizing has, at best, an equilibrium. The complexity ladder tracks it — solving an MDP is tractable (P-complete), the finite-horizon single-agent POMDP is PSPACE-complete, and the finite-horizon multi-agent case is NEXP-complete even in its cooperative form, the decentralized POMDP (Bernstein et al., 2002) — provably beyond polynomial time, and far harder than NP. No one solves this object directly; the working models are its tractable reductions — single-agent impact control (Almgren & Chriss, 2000) and, as participants multiply, its mean-field-game limit (Lasry & Lions, 2007). Optimality is off the table; an edge is not. ↩
-
Pedro Domingos (@pmddomingos), on X, April 24, 2026. ↩
-
One operator's probe, not a benchmark — but it tracks the broader frontier. In 2025, AI systems reached gold-medal standard at the International Mathematical Olympiad, graded under the same conditions as human competitors (DeepMind's Gemini Deep Think, with a comparable OpenAI result); and, working alongside mathematicians including Terence Tao, models have begun producing genuinely new results — improved bounds and novel constructions on long-studied problems — under careful human verification (DeepMind's AlphaEvolve, 2025). ↩