Much of the current AI discussion in trading still centres on agents, prompts and models, with questions continuing to be asked about how AI will work in practice. Meanwhile, a deeper shift is already underway toward engineering the infrastructure that allows AI to operate inside real trading systems where governance, latency and execution constraints will matter as much as model quality. The challenge in trading is not discovering a signal; it is operationalising that signal across market data pipelines, execution systems, governance frameworks and latency-sensitive infrastructure – market connectivity.

AI adoption in trading is increasingly developing across three layers:

• Decision support – AI interacting with market data and research workflows
• Execution infrastructure – systems operationalising AI-generated strategies and
• Market rails – infrastructure enabling automated interaction with markets

Here’s what I learnt this week on AI in Trading:

1. AI Research Systems Are Moving Closer to Trading Workflows

A new collaboration between NVIDIA and KX, with a proof-of-concept deployment at RBC Capital Markets, introduced AI blueprints for research and signal generation. The architecture combines GPU-accelerated vector search, real-time time-series analytics and multimodal retrieval to analyse market data, research, filings and news simultaneously. The RBC pilot reportedly compressed elements of the research workflow from hours to minutes. One design feature is the use of “temporal AI,” aligning all inputs to market event time so signals can be evaluated in a point-in-time correct context – read more here: https://kx.com/news-room/kx-launches-agentic-ai-blueprints-powered-by-nvidia-at-gtc-2026-featuring-a-capital-markets-research-assistant-and-trading-signal-agent/

Why this matters for Trading: The real breakthrough is not the AI agents themselves but the infrastructure connecting AI to real-time market data and trading workflows. As research automation accelerates, firms that can embed AI insights directly into live trading environments without sacrificing latency, governance or operational control will gain the advantage. How this changes the relationship between portfolio managers and trading desks will be one to watch as research gets ever closer to execution.

2. From Single Agents to Multi-Agent Swarms

AI development in trading also continues to move beyond single-model forecasting toward multi-agent systems. DeepAgents from LangChain demonstrate how AI agents can plan tasks, coordinate workflows and execute complex processes rather than simply responding to prompts – read more here: https://docs.langchain.com/oss/python/deepagents/overview.

Another project gaining attention is MiroFish, which models complex environments using thousands of interacting AI agents. Instead of generating a single prediction, the system builds a simulated environment where agents representing investors, policymakers and information flows interact, allowing users to explore how events propagate across markets. Read more here:
https://github.com/666ghj/MiroFish/blob/main/README-EN.md

Why this matters for Trading: Simulation-based modelling is emerging as a powerful decision-support tool for understanding how signals and information propagate through markets. Unlike traditional machine-learning models, which analyse historical data to generate statistical forecasts such as price direction or volatility, these systems create simulated environments populated by thousands of interacting agents. By introducing events such as news, earnings surprises or policy changes, firms can observe how behaviour evolves across the simulated market. Rather than producing a single prediction, this approach generates multiple potential scenarios, allowing trading teams to stress-test strategies and better understand how sentiment and information may spread through complex market systems before deploying capital.

3. From Single Prompts to Systems Engineering

As AI systems evolve from single models to multi-agent environments, the next challenge becomes coordinating and supervising those agents in production. New infrastructure tools are emerging to support this transition, including the 10XT Control Plane from 10XTraders.AI, which aims to operationalise AI trading strategies across monitoring, orchestration and deployment – retrieval-augmented generation grounded in proprietary market data, multi-agent orchestration across workflows, persistent state and memory systems, model evaluation and monitoring frameworks – as well as human-in-the-loop governance – read more here:
https://www.globenewswire.com/news-release/2026/03/15/3255918/0/en/10XTraders-AI-Introduces-the-10XT-Control-Plane-Cloud-Native-Infrastructure-for-AI-Trading-Systems.html

Why this matters for Trading: These developments reflect a broader shift from experimenting with standalone AI models to building structured systems capable of running multiple AI components inside live trading environments. As AI begins to operate across research, strategy and execution workflows, the challenge becomes operational reliability rather than model success. The competitive advantage is increasingly moving towards AI infrastructure and systems engineering.

4. Meeting the Need for Governance

A Reuters report on 12 March highlighted that EU economies are backing stronger centralised supervision of European capital markets – read more here: https://www.reuters.com/business/finance/germany-five-other-biggest-eu-economies-back-centralised-market-supervision-2026-03-12/. The proposals include greater oversight across trading venues, crypto-asset providers and clearing houses as markets become increasingly digital and cross-border. Regulatory attention is shifting toward the infrastructure layer where automated trading systems interact with markets and the need to treat financial data architecture as critical public infrastructure to enable consistent, cross‑sector and cross‑border oversight – read more here from Guilia Ferris @ ESMA –  https://lnkd.in/dSgvJ9pG.

At the same time, independent technology benchmarks are increasingly being used to evaluate the infrastructure supporting AI workloads in finance. Recent Strategic Technology Analysis Center (STAC-AI) benchmark results showed significant performance improvements in financial LLM inference on NVIDIA’s Blackwell architecture, using workloads based on analysing EDGAR filings – read more here: https://www.mexc.com/news/862278

Why this matters for Trading: AI adoption in trading increasingly depends on governance, explainability and operational control – not just model capability. As regulators focus more closely on market infrastructure, firms will need to demonstrate that AI systems operate on tested, transparent and auditable technology stacks. Independent benchmarking initiatives such as STAC-AI help provide evidence of system performance and reliability under real financial workloads, supporting technology due diligence and operational risk management.

5. One Step Closer to Tokenised Markets

Another major development this week was the announcement of a partnership between Nasdaq and Kraken to explore tokenised versions of listed equities. The initiative focuses on modernising post-trade infrastructure through faster settlement, automated corporate actions and broader distribution. While a potential launch is not expected until 2027 – and tokenisation is not AI itself – the development reinforces the trend toward more programmable market infrastructure, allowing software systems to interact directly with trading and settlement processes – read more here: https://www.reuters.com/business/nasdaq-teams-up-with-kraken-expand-tokenization-infrastructure-2026-03-09/

Why this matters for Trading: More programmable markets make trading infrastructure increasingly machine-compatible. As markets evolve toward longer trading hours and potentially continuous 24/7 trading – AI systems will play a growing role in monitoring markets, executing strategies and providing liquidity. This has significant implications both for market infrastructure and for the traders who operate within it, making disciplined systems engineering and operational controls more important than ever.

Markets are not becoming autonomous overnight, but they are becoming increasingly engineered, data-driven and compute-intensive. As with algorithmic trading before it, once a small number of firms demonstrate clear performance advantages, adoption is likely to accelerate rapidly across the industry. What felt inconceivable less than two years ago is now moving firmly into production – AI in trading is no longer a theory, it is becoming part of market infrastructure. The question is not whether AI will be used in trading, but how quickly firms can build the systems, governance and expertise needed to deploy it responsibly and competitively. One to keep watching.

As always, thank you for reading – any comments or feedback as always welcome.

Rebecca