Agentic AI Now Runs Research: Public Scoreboard Shows Why Humans Still Run the Money

Automation in trading has quietly changed shape.

For most of the last two decades, automation meant execution. Slice a large order, hide your intent, get a better fill. That argument is over. Algorithms accounted for 50% of NSE cash market turnover for the first time in FY25, up from 17% in FY11. In equity derivatives, algo share is closer to 70% of notional turnover (NSE Market Pulse data, via Business Standard).

What has changed recently is that automation moved one layer up, into research itself.Agentic AI trading systems now reason, use tools and work in coordinated teams. The era of asking a chatbot for a few lines of Python and then debugging them for an hour is largely behind us.

But something else has not changed at all. Every serious attempt to hand an AI model money and let it trade on its own has gone badly. That gap between AI as a research engine and AI as a decision-maker is the real story.

Where the ideas were dying

Most quantitative researchers have a notebook full of ideas they never tested.

The reason is rarely a shortage of ideas. It is the work in between. Turning a plain-English hypothesis into a clean, trustworthy backtest takes hours or days: pulling data, checking it, writing the simulation, fixing the bugs, then discovering a subtle look-ahead error and starting again.

The common workaround has been to treat a general AI model as a one-person team. Ask it to be a data scientist, coder and risk manager at once. In practice, that produces a loop of error-fixing and re-prompting rather than research.

Agentic systems take a different route. Instead of one generalist, you get a small team of specialists.

The architecture of the virtual research desk

In a modern systematic setup, automation is no longer one script on a schedule. It is several agents, each owning one part of the workflow:

• The Hypothesis Designer turns a vague idea, such as "buy when the market looks oversold," into a precise, testable specification: which assets, which timeframe, what entry and exit logic.
• The Data Scout fetches historical data and computes the indicators, and checks the data is clean enough to simulate on.
• The Backtester writes the code that simulates the buy and sell decisions on historical data.
• The Performance Analyst measures the result: risk-adjusted returns, and maximum drawdown, meaning the largest peak-to-trough fall the strategy went through.
• The Assembler stitches everything into a research document, so a raw idea becomes a documented experiment.

This is not hypothetical. Man Group has published details of AlphaGPT, its own agentic research workflow, which it describes as a "digital three-person research team": an Idea Person that generates hypotheses, an Implementer that writes production-grade Python against internal databases, and an Evaluator that applies statistical significance tests, risk checks and economic reasoning checks. An orchestrator coordinates them and adds safeguards against hallucination (Man Group, November 2025).

The firm reports the system produced dozens of viable concepts in minutes, and explored relationships its human researchers had not covered. Not because those researchers lacked skill, but because there are only so many hours in a week.

The speed gains show up elsewhere too. Balyasny Asset Management says roughly 95% of its investment teams use its internal AI research platform, and that one agent cut a macroeconomic scenario analysis from two days to about 30 minutes (OpenAI case study, March 2026). These are self-reported figures in a vendor case study, so treat them as directional. But the direction is consistent across firms.

What the public tests actually showed

Now the other side.

In 2025 and 2026, a startup called Nof1 ran a series of contests called Alpha Arena. Eight frontier AI models, including Claude, Gemini, ChatGPT and Grok, were each given $10,000 and set loose on US tech stocks for two weeks at a time, across four separate challenges.

The portfolio as a whole lost about a third of its capital. Across all 32 sets of results, a model finished in profit only six times. A separate tracker, the AI blog Flat Circle, followed 11 markets-related arenas. In only two was the median model profitable (Bloomberg, via Financial Advisor Magazine, May 2026).

These contests are short, small and not academically rigorous. They do not prove that agentic trading cannot work. But one finding is hard to wave away.

Given identical instructions, Grok placed 158 trades. Qwen placed 1,418. One model preferred to go long, another was happy to short, another took large leverage. As Doug Clinton, who runs an LLM-driven fund, put it, the models "have personalities that you have to manage almost like a human analyst."

A system whose behaviour varies that widely under the same instruction cannot be given size, whatever its average return looks like.

Nof1's founder Jay Azhang was direct about it: "LLMs can't really make money by themselves. You need basically a very sophisticated harness and scaffolding and data platform to even give them a chance."

That harness is the actual work.

Four risks to design around

The ease of generating strategies creates methodological problems. Test enough random ideas and some will look profitable purely by chance. Anyone usingAI in trading research needs to manage four risks.

1. Hallucination. Models do not know when they are wrong. They produce fabricated figures, invented references, or strategy logic that reads well but contains an error only a domain expert would catch. Man Group also flags a related failure it calls drift: the AI conceptualises one research idea and implements a slightly different one.

2. Data snooping bias. The more ideas you test, the more likely you are to find a lucky result rather than a real edge. This is not new. Harvey, Liu and Zhu argued back in 2016 that given the scale of factor mining in finance, a newly discovered factor should clear a t-ratio above 3.0, not the conventional 2.0 (Review of Financial Studies, 2016). Agentic AI did not create this problem. It multiplied an existing one.

3. Overfitting. A model can fit history beautifully and still be useless on tomorrow's market.

4. Lookahead bias in evaluating the agent itself. This one is specific to LLMs and is often missed. Ask a model in 2026 how it would have traded March 2020, and it already knows how March 2020 ended. The default validation tool in quantitative finance, running a strategy backwards through history, does not work cleanly on an agent that has read the history. This is precisely why live arenas exist. The practical implication: use agents to generate and code hypotheses, but do not let an agent that has memorised the period be the judge of whether its own idea worked in it.

The response from experienced desks is not to test less. It is to raise the bar. Stricter statistical thresholds. More realistic transaction cost modelling. The more realistically a backtest simulates each trade, the harder it is for a weak strategy to pass, which is the point. And an explicit audit layer, because AI-generated backtests should never be assumed correct.

Man Group's guardrails are a reasonable template. Every assumption and decision step logged and reviewable. Dual-track validation, with an Investment Committee reviewing the economic rationale and technology teams separately reviewing the code. And every AI-generated signal required to clear the same thresholds as a human-generated one.

Democratisation, with an asterisk

Agentic AI does level part of the playing field. Processing alternative data such as satellite imagery, news sentiment or transaction data used to require a data engineering team. An independent trader can now build an agentic workflow to turn unstructured sources into signals. Large firms such as Renaissance Technologies and Two Sigma have run machine learning on alternative data for years. The tooling to attempt something similar is now within reach of a much smaller operation.

But the asterisk matters. Institutional spending on alternative data is going up, not down. In a survey of senior staff at firms managing around $2.4 trillion, 84% reported spending between $500,000 and $2.5 million a year on alternative datasets, and 54% said those budgets had grown by 50% or more over two years (Hedgeweek).

So what has been democratised is the tooling to process messy data, not access to the best data itself. And the cognitive barrier has not moved at all. Cheap access to data is worth little without the skill to interpret it and the discipline to reject most of what it suggests.

In India, the regulatory layer has caught up with the technology. SEBI's framework on retail participation in algorithmic trading, first issued in February 2025 and mandatory from 1 April 2026, requires every algo order to carry an exchange-issued Algo ID so trades can be traced. Brokers act as principal and are accountable for the algos running through their platforms, and algo providers must work through a registered broker rather than connecting directly to exchanges (SEBI).

The context behind that framework is worth knowing. SEBI's own study found net losses of individual traders in equity derivatives rose 41% year on year to about ₹1.06 lakh crore in FY25, with roughly 91% of individual traders losing money (Business Standard, July 2025). That figure is about derivatives trading generally, not algos specifically. But it explains why more automation is being permitted with traceability attached rather than without it.

What this means for the modern quant

If you already run a systematic process, the sensible path is narrow and unexciting. Start from an edge you have already validated, and use agents to explore around it faster.

Shift from coding to orchestrating. Your job moves from writing every line to designing the system, the tool access and the guardrails the agents work inside.

Build the audit layer first, not last. Assume AI-generated backtests contain errors until checked. Automate the checks for the common ones: look-ahead leakage, survivorship bias, unrealistic fills, missing costs.

Use specialised setups over general ones. Systems built around a specific market or asset class, with the right data and the right tools attached, tend to be more reliable than a general model asked to do everything.

Insist on the economic reason. A statistical pattern with no explanation behind it gives you no way to judge whether today's market still supports it. Man Group requires every AI-produced signal to demonstrate a clear economic rationale before it can be considered. That is a good standard to borrow.

The most effective workflows combine automation with oversight. AI can compress research and handle execution. Decisions about strategy design, validation, position sizing and capital allocation still sit with a person. You decide how much capital the system gets, what it rejects, and when it stops.

As with any trading, systematic or discretionary, losses are possible and no approach performs well in all market conditions.

The edge in agentic AI trading will not go to whoever runs the most powerful model. Man Group expects the technology itself to become widely available and says its own advantage lies in research philosophy and methodology refined over market cycles, not in early access to a model. The edge goes to whoever asks better questions, validates the answers more carefully, and knows when to override the machine.

The tools are becoming widely available. The judgment to use them well is not.