A symphony of agents orchestrating smarter stock trading

In 2025 we are observing how agentic AI is changing the way financial markets work. It is no longer a passive tool but an autonomous partner in decision-making.

BlackRock’s Asimov system scans filings, reports, and internal notes to provide instant insights to portfolio managers, while AllianceBernstein applies AI to compress months of research into hours.

Regulators are raising concerns, with the UK’s FCA warning that AI is advancing faster than the rules governing markets. This signals a structural shift as agentic AI moves beyond trade execution and brings speed, context, and strategy into investment decisions.

Adoption of artificial intelligence (AI) and agentic AI in stockmarket trading has seen multifold growth in recent years. It is expected to increase from $21.59bn in 2024 to $24.53bn in 2025, showing a compound annual growth rate (CAGR) of 13.6 per cent, according to The Business Research Company report 2025.

Traditional algorithmic trading relies on data-driven models that support human decisions but struggle with real-time market shifts. Agentic AI introduces autonomous decision-making and self-learning capabilities. This shift helps financial models to adapt with dynamic changes. It also helps in optimising strategies that require minimum human interventions.

Agentic AI empowers financial systems to autonomously trade, detect fraud, and manage risk by applying advanced learning methods. It can continuously adapt decision-making processes in real time. Mostly it uses techniques like neural-symbolic learning and Bayesian inference to process the requests.

Agents in harmony

Trading is no longer a solo journey. It now runs on multi-agent systems working in tandem. Specialised AI agents manage market feeds, technical analysis, sentiment, scalping, and back testing simultaneously. The Market Data Agent tracks price ticks, volumes, spreads, volatility, and correlations. The Technical Analysis Agent applies indicators, liquidity zones, and historical patterns to generate signals. Together, they power an adaptive, intelligence-driven trading ecosystem.

But it is necessary for all parts of the orchestra to be playing in a well-rehearsed symphony, rather than producing a cacophony of discordant trades.

Market sentiment analysis is an AI technique that leverages integration with large language models (LLMs) like FinBERT, and domain-specific natural language processing (NLP) models to analyse public sentiment from earnings reports, financial news, and social media trends. By leveraging contextual embeddings and reinforcement learning, AI models execute sentiment-driven trades with higher precision.

Risk manager agents help to reduce risk by monitoring the risk reward ratio for a trade. Through VaR (value at risk), it can calculate how much one could take a loss on a bad day. It requires users to define maximum portfolio dropdown limits.

Agentic AI is fast reshaping regulatory compliance. It enables automated reporting, continuous monitoring and stronger risk governance. In high-frequency trading, this monitoring is critical. It ensures trading activity stays within regulatory limits and prevents market manipulation.

All of this is overseen by the orchestra conductor: the LLM inference engine is the command centre of multi-agent trading. It combines reasoning with action to design strategies and execute them through external tools. It’s self-reflection loop reviews past trades, corrects flawed logic, and improves future decisions. Instead of direct buy or sell calls, it generates scores that downstream systems translate into trading signals.

In summary, the above steps can be broadly divided into three major categories in business parlance.

Data preparation allows AI agents such as market data agent, technical analysis agent, or portfolio agent to generate stock-related data. This helps in deriving critical insights.

These collected data are then fed into LLM models, which take clean and prepared label data and produce a knowledge base for the LLM model. This step leverages a graph neural network (GNN) to generate embeddings that can be then fed into XGBoost to produce the stock model for better strategy and execution.

But the most important part is played by the LLM inference server, which will take the prepared data from the knowledge base to come up with a score. This score will be considered to come up with buy/sell signals later. The inference server never decides on the final buy/sell recommendations, but only produces a score.

Financial markets are rapidly adopting agentic AI. Success requires a multidisciplinary approach that combines innovation, ethics, and regulatory foresight. Technology alone is not enough.

A sound trading strategy is critical. Without it, outcomes can be damaging. In conclusion, it is worth recalling the famous words of Arthur C. Clarke, which remind us that “any sufficiently advanced technology is indistinguishable from magic”. The time has come for the industry to align its processes, in order to realise the full potential of agentic AI in stock trading. Every beautiful musical piece must have a distinctive coda.

Soumya Bhattacharya is a lead enterprise architect at IBM  

Raja Basu is a financial market leader, SME at IBM