AI moves further into the execution stack with Agents for All
OpenClaw has accelerated the shift to AI agents by showing that value no longer resides in the model alone, but in the surrounding architecture that enables autonomous deployment across systems. This reframes AI from an analytical tool into an operational control layer. With that shift comes a new class of risk: not just how models perform, but what is connected into the control layer, where, and with what permissions, creating vulnerabilities akin to supply chain attacks. Competitive advantage therefore moves to how firms design and govern their agent stack. In trading, this signals a transition from standalone tools to re-architected workflows – raising a critical question: while anyone can now build an agent, who can do so safely and at scale? Here’s what I learned this week on AI in Trading:
1. Agents go mainstream: GameStop 2.0?
Public has introduced AI-powered brokerage agents that automate trading, allocation, hedging, and cash management – effectively putting retail portfolios on “autopilot” with transparency and user control. By enabling custom indices and more complex strategies, it lowers the barrier to institutional-style investing. Read more here – https://www.wsj.com/tech/ai/buying-the-dip-this-ai-agent-will-do-it-for-you-1d2b1658
What this means for trading: If automated retail agents become a structural driver of trading workflows, this shifts retail flow from discretionary to systematic execution. While broker disintermediation is unlikely in the short term – that will require generational change – retail flow could become more predictable, time-compressed, and correlated – concentrating liquidity demand into windows such as the open, close, and rebalancing events – just as these YouTube videos suggest:
But concentration also raise price impact and reduces resilience. Synchronized, rules-based strategies amplify flow clustering and feedback loops, increasing the risk of intraday volatility and dislocations. As retail becomes a systematic component of liquidity formation, markets grow more susceptible to mechanically driven moves and herd dynamics. This is reinforced by the rise of 0DTE options – now citing >60% of S&P 500 (SPX) options volume per Cboe – which further shifts price formation toward positioning and hedging flows on short term time horizons rather than company fundamentals (https://ir.cboe.com/news/news-details/2026/Cboe-Global-Markets-Reports-Trading-Volume-for-February-2026/default.aspx).
2. Retail agents meet institutional complexity
While retail agents may be about to start in earnest – those who are building agents in enterprise are warning of the growing complexity when moving from “single autonomous models” towards multi-step (MAS swarms), orchestrated systems embedded in workflows. Agents need to be built as tool-using components within structured pipelines – combining data ingestion, reasoning, strategy generation, and execution layers, with explicit control, validation, and human oversight (read more here – https://arxiv.org/abs/2603.13942).
What this means for trading: As agentic trading is moving beyond models toward embedding systematic execution at scale, the constraint is not intelligence but architecture – specifically orchestration, interoperability, and observability. This requires decomposed step by step multi-agent workflows with guardrails, cost controls, and human-in-the-loop design. As retail agents converge on similar signals and time horizons, their flow will increasingly resemble institutional systematic flow. The key variable then becomes how agents are distributed, coupled, and governed: poorly constrained or highly correlated behaviour increases the risk of crowding and self-reinforcing volatility. Read more here – https://www.techradar.com/pro/agentic-ai-transforming-industries-and-tackling-the-interoperability-imperative)
3. Case for caution: Claude Code leak
Anthropic’s accidental exposure of Claude Code highlighted how modern AI systems are built around orchestration layers, memory, and multi-agent coordination. While no model weights or user data were leaked, the incident revealed exactly how architectural dependencies and emerging risks are arising around always-on agent workflows. Read more here – https://www.axios.com/2026/03/31/anthropic-leaked-source-code-ai
What this means for trading: If the key vulnerability lies in the control layer – this shifts the focus of risk and governance to the execution stack, where agents operate across routing, allocation, and infrastructure. While this expands the attack surface, there is an argument that this could enable more robust operational governance if agent workflows are structured, observable, and continuously monitored. This would allow firms to embed controls directly into the decision-to-execution lifecycle through granular permissioning, real-time validation, and systematic stress testing. As a result, governance becomes programmable and proactive rather than reactive, potentially leading to more resilient and controllable market infrastructure despite the increased complexity.
4. JPMorgan TradeFM: modelling markets as events
JPMorgan’s TradeFM is a 524M-parameter generative model that predicts the next market event – timing, size, depth, and direction – rather than price, marking a shift toward modelling market microstructure as a sequence of events. Trained on billions of trades across more than 9,000 equities, it uses scale-invariant features and a unified tokenisation approach to normalise trading data, enabling cross-asset and cross-market generalisation without asset-specific calibration. Integrated with a market simulator, it reproduces key market dynamics such as volatility clustering and heavy tails, while delivering 2–3x lower error than traditional Hawkes-based models and generalising effectively to out-of-sample markets like APAC. The results suggest that transferable structure exists in order flow, opening applications in synthetic data generation, stress testing, and learning-based trading systems. Read more here – https://arxiv.org/abs/2602.23784.
What this means for trading: This signals a shift from price prediction to modelling market behaviour as event sequences. The edge moves toward understanding liquidity formation and flow dynamics, potentially improving execution, simulation, and risk management across markets making the competitive advantage in understanding and anticipating flow dynamics rather than forecasting price direction alone.
5. Tokenisation, “FutureFi,” and systemic risk
The convergence of AI agents and tokenised markets is creating always-on trading environments where liquidity appears continuous, but correlations intensify and boundaries blur. However as highlighted at RegTech2026 and FIS Singapore, while automation is shifting – accountability is not: despite increasing autonomy, responsibility for outcomes remains firmly with the CIO or asset owner. Read more here – https://www.top1000funds.com/2026/04/ai-will-revolutionise-investing-but-machines-wont-carry-the-can
What this means for trading: AI-driven execution introduces risks around bias, opacity, and cyber vulnerability. Combine AI with stablecoins and on-chain infrastructure, this creates new failure modes – asset risk, plumbing risk, and run dynamics. These systems are “run-optimised”: always-on, global, and frictionless. Paired with AI, this increases the risk of rapid, synchronised dislocations – shifting the need of regulatory focus from execution towards governance, control, and operational resilience (read more here https://www.suerf.org/publications/suerf-policy-notes-and-briefs/stablecoins-are-run-optimised-instruments/).
Infrastructure constraint: compute, energy, geography
The ability to implement AI in Trading though still depends on the basics – data, data centeres and energy. New research from Cambridge highlights the impact of the growing increase of AI-based services and the resulting heat dissipation of AI hyperscalers. The paper estimates the land surface temperature increases on average by 2°C but potentially 9°C creating a “data heat island” effect which could impact 340 million people (read more here – https://arxiv.org/abs/2603.20897).
Nebius Group’s planned 310 MW data centre in Finland – backed by multi-billion-dollar agreements with Microsoft and Meta highlights how cold climates and stable energy grids are now strategic assets for compute. Read more here – https://www.reuters.com/technology/nebius-furthers-european-expansion-with-10-billion-ai-data-centre-finland-2026-03-31/
AI strategy is now inseparable from energy, geography, and physical resilience; competitive advantage is not just moving from models to orchestration but chip access, power, siting, and infrastructure. Even grid stability and zoning approvals are becoming as strategically important as cloud architecture. With the latest in Iran – one to keep watching.
Thanks again for reading – more to follow next week. As always let me know your feedback/comments
Many thanks
Rebecca
