According to Jon Light, Senior Director of Product Management at Devexperts, “agentic brokerage” is becoming a loaded term in retail – but the genuinely interesting shift is structural, not cosmetic. Most AI in brokerage today is still assistive: chatbots, market summaries, onboarding tools that respond when prompted and then go quiet. An agentic system stops being reactive and starts being persistent and action-oriented, monitoring, reasoning and acting within defined permissions and guardrails. Light frames it as the third in a sequence of platform cycles: the internet produced the commission-free broker, the smartphone produced the mobile-first feed-driven broker, and AI is now producing the agentic broker – platforms that stop being interfaces to the market and start operating as continuous intelligent layers across the stack.
The trader-side benefits are the obvious entry point – less noise, more personalisation, structured intelligence rather than a constant stream of data – but Light argues the real structural impact lands on the broker. With average retail relationships short-lived, agentic systems shift focus from acquisition to retention, lifetime value and revenue stability by detecting early behavioural changes – declining engagement, shifts in risk appetite, deviations from a trader’s usual patterns – in time to intervene rather than retrospectively. The most significant change, he says, is in risk management: today’s broker risk systems are largely reactive and aggregated, looking at exposures after they have built up. Continuous, real-time behavioural risk intelligence allows brokers to see leverage, revenge trading or stress signals at the individual trader level, moving the discipline from post-event analysis to pre-event intervention.
A common assumption is that bolting an LLM onto a trading platform is enough to make it agentic. Light is direct that it is not. LLMs handle language, reasoning, summarisation and orchestration well, but they cannot model markets, model risk or capture behavioural dynamics at scale – that work belongs to machine learning models trained on structured trading data, handling churn prediction, risk scoring, behavioural clustering, profitability forecasting and time-series pattern detection. The agent becomes the layer that coordinates between them. Underpinning all of it, he stresses, is API quality and the emerging Model Context Protocol layer that sits on top: without a clean, structured, accessible surface for agents to interact with the broker’s tools, data and actions, the agentic brokerage becomes fragile and fragmented. The future, on Light’s reading, is not a single large model running inside a brokerage – it is a system of specialised models and agents working together across a strong, governed API layer.
Subscribe to our newsletter
