The Bank of England and FCA survey found 80% of UK firms running agents but only 2% granting broad autonomy. Jefferies mapped US brokers opening up to agents, Robinhood reported around 200,000 agentic accounts, and CFTC Chairman Michael Selig opened the Frontier Forum asking whether markets now need to “know-your-agent” (KYA). Tradeweb’s Q3 call showed AI investment accelerating, while the ICI set out how AI should prevent operational breaks, and Quod Financial and Beeks launched tools to ensure Gen AI is kept away from the order.

Agent adoption is increasingly tiered – not just fast in the US and slow in Europe. What matters is who decides what a system does next: an algo sets order parameters inside a strategy humans fixed in advance; whereas an agent decides at runtime which tasks to do, with which tools and data. Here are the five things I learnt this week on AI in Trading:

1. The BoE and FCA survey: beyond the headlines

The survey covers 196 firms but only 100 from investment and capital markets:

·        Decision support dominates – 80% run agents that recommend while humans decide, but 46% allow constrained execution within limits.

•       Trading is the most restricted – 7% use decision-support agents in trade and execution – yet just 14% use AI-driven (ML)execution: most trading AI is still algos. This stat alone is likely higher among trading firms.

•       Vendors hold the technology – third-party systems are 39% of AI systems and over half the high-materiality ones; 81% rely on vendor assurance, while only 24% claim deep knowledge.

•       Controls are front-loaded – over 90% review before deployment; around half monitor afterwards and only 49% have AI incident playbooks.

•       Concentration is rising – the top three cloud providers hold 79% of relationships, the top three model providers 54%; and just a third of critical cloud relationships have exit plans. Firms expect interconnectedness to be the fastest-growing systemic risk, ahead of AI-enabled manipulation. Firms now need to evidence what a vendor system does today and whether it still delivers within risk appetite, as the FCA’s good and poor practices also stress.

Why this matters for trading:

•       The theoretical blocker has moved – from safety, talent and explainability in 2024 to integration (23%) and legacy systems (22%) now. Making agents reliable and accountable inside real processes is an operating-model problem, not a technology one.

•       Decision support is still agentic. On the OECD definitions the survey uses, an agent that calls tools and drafts tickets is an AI agent – agentic AI co-ordinates several – and needs identity, scoped permissions and trajectory logs, not just a human at the end.

•       “Human decides”: approving 200 agent tickets an hour is arguably “limited human intervention”, where RTS 6 applies again – and a sign-off that cannot catch microscopic drift only decides who takes the blame.

•       Observability makes agents auditable – what an agent saw, which tools it called, what it recommended and who acted. Apply the same test to broker algos, EMS analytics and TCA.

•       “The vendor said so” is still the norm – the FCA’s good and poor practice flagged untested vendor assurance, and a quarter of a small sample let vendor model updates go straight into production.

2. Washington embraces SI: who is the agent trading for, and is it safe?

At the Frontier Forum, Selig said “super intelligence promises to reshape our financial markets” (an awkward “SI” for EU readers) and asked:

•       Registration – does an agent’s developer need to register?

•       Liability – who is liable if the agent goes rogue?

•       Identity – is it time for “know-your-agent” (KYA), as we have advocated in the FIX AI Steering Group?

•       Market risk – could correlated models and autonomous speed stretch surveillance?

He proposed no new rules. Architect’s Brett Harrison described AI putting black-box algorithms in far more hands; the ICI’s Andrew Kayiira and OneChronos’ Kelly Littlepage were among the panellists, and a 24/7 roundtable is reported for 15th December. The questions converge with the FCA’s trusted framework of agent identity, delegated authority and accountability.

Why this matters for trading:

•       KYA is the missing order field – nothing flags agent-originated orders in FIX or transaction reporting, so no desk knows whether it trades against a person, an algo or an agent. The FIX AI Working Group’s work with GLEIF’s verifiable LEI could supply it.

•       Responsibility splits three ways – the client decides, the AI platform chooses the trade outside the perimeter, and the broker carries the controls (RTS 6, Rule 15c3-5). Registering developers would redraw that map.

•       Correlated models are a US concern too – per-algo limits don’t aggregate across agents sharing a provider or signal.

•       24/7 is next – prices increasingly form at weekends, as the FCA’s Nikhil Rathi noted; that is a European question too.

3. US retail agents pass 200,000 accounts

Robinhood’s Q3 results showed around 200,000 agentic accounts and over $1.6bn traded; Vlad Tenev said they hold over $250m. Modest against close to $400bn, but each is a deliberate decision to hand trading to an agent. Others:

• Interactive Brokers – autonomous trading since 21 October in AI Partitions across 75,000+ stocks, options and ETFs, with any MCP model (release).
• tastytrade – an official MCP server for equities, options, futures and crypto, with dry-run checks on every order (docs).
• Public – agents trading portfolios since March, prediction markets since September (launch).

US banks go the other way: Axios reported banks blocking unvetted agents in customer accounts, JPMorgan won’t let outside agents in and Citi routes them through its own platform. Access is now closed (banks), permissioned (own platform only), partitioned (IBKR, Robinhood: funded, capped accounts) or advisory (UK and European institutions). Europe is not insulated: Scalable Capital opened to ChatGPT and Claude this summer, and US agents trade the same global stocks around the clock.

Retail flow is also an institutional signal. Citadel Securities, which executes roughly 35% of US retail volume, showcased Breakout Probability at its Future of Global Markets conference: daily squeeze odds for US stocks from retail stock and options flow plus short interest, backed by 500+ quant researchers and 200+ petabytes of data. Its example was Avis Budget: up almost 390% in a month in April, with 86% of the float short, two funds holding over 71% and shorts losing over $4bn, before collapsing within days (Quartz).

Why this matters for trading:

•       Most European desks will first meet autonomous agents on the other side of their trades – same-model mandates firing together.

•       Tiered access isn’t tiered exposure – when AI stocks sold off this week, nobody could say how much of the selling came from agents.

•       Retail herding is invisible to any single broker – but not to the wholesaler, now selling squeeze forecasts: a timing risk your data cannot see.

4. Increasingly the vendor builds the trading tools – faster

Rather than waiting on brokers to roll out tech – vendor partnerships are increasingly delivering. Quod Financial launched Qanvas: desks can request OMS and EMS features in plain language (one took four hours 33 minutes) and AI builds the front end. Beeks and OmniPhi announced a sovereign agentic platform on privately hosted models (London and New York, Q1 2027) that cut Beeks’ own algo development from nine months to days. Tradeweb’s Q3 call flagged an upgraded AI Price engine and $4.5m of extra AI cost.

Why this matters for trading:

•       “Plan at set-up, run deterministically” keeps systems in line with algos and current RTS 6. Like LTX last week, Quod is agentic at build time and deterministic at runtime.

•       Agents writing strategy code is development – faster build times now make algo change control the new bottleneck.

•       Natural language multiplies strategies but could also impact testing – each plain-English configuration may be a new strategy under ESMA’s “testable and distinguishable” test.

•       Venue AI will increasingly shape price discovery – evaluated prices like AI Price feed dealer selection and TCA.

5. The ICI: use AI to stop the break, not fix it faster

The ICI’s Financial Innovation Advisory Committee published an “Intelligent Operating Model”. Noting  T+1 and extended hours shrinking the time to fix problems, its examples include catching settlement issues before trade details leave the building and linking redemptions to liquidity decisions.

Why this matters for trading:

•       No orders, still agents – operations agents need identity, least-privilege access and logs too; allocations, SSIs, confirmations and fails are high-volume and measurable – a useful start for agent implementation.

•       T+1 sets a deadline – Europe moves on 11th October 2027, and is focusing more on automation of process rather than T+1 as the sole objective. AI is only as good as the data joins beneath it and as markets move to T+0, this will become increasingly evident.

This week saw regulators continue to converge on one question – who is the agent, and whom does it act for – and vendors on one design: AI builds the tools but still relies on deterministic controls for execution. But algo versus agent is not safe versus unsafe: gateway controls, price collars, throttles and the kill switch remain the floor; but on top will be the controls never required for an algo operating in isolation – identity per agent, least-privilege tools, trajectory logs, vendor observability, red-teaming and a checkpoint between “agent proposes” and “action executes”. European caution can keep its own agents advisory, but that does not include the rest of the market. The question for every Head of Trading is who, or what, is on the other side of your trade – and whether you would even know.

Thanks for reading as always.

Rebecca