As discussed in last week’s newsletter, agent-led execution continues to move from experimentation into core trading workflows. Machine-generated signals are now embedded directly into execution, risk, and surveillance frameworks – reshaping market microstructure in the process – and now the regulators are catching up and letting the industry know their expectations. Here’s what I learnt this week on AI in Trading:

1. UK: From Innovation to Resilience

The UK has focussed AI engagement on the FCA’s sandbox and innovation initiatives – with the latest invitation here – https://www.fca.org.uk/news/news-stories/applications-open-second-cohort-ai-live-testing. However, this week the UK Treasury Select Committee published a report concluding that in their opinion – the FCA, Bank of England (PRA), and HM Treasury are not moving fast enough to manage the risks created by widespread AI adoption in financial services – read more here – .

If AI is no longer experimental and is embedded across algorithmic execution, market surveillance, and risk modelling, that shifts the risk profile according to the Committee – highlighting particular concern around opaque “black box” models, governance and senior manager accountability, and regulatory compliance under SM&CR. It also warns that common models or shared data sources could amplify volatility and pro-cyclical behaviour during periods of stress, while operational and cyber risks are heightened by concentration in a small number of AI and cloud providers. Baseline themes which have been highlighted previously going back to the Foresight report (https://www.gov.uk/government/collections/future-of-computer-trading). However, the expected future regulatory direction is towards greater scrutiny rather than restriction, not to slow innovation, but to ensure resilience: with system-wide AI stress testing, stronger expectations on explainability, clearer accountability when automated decisions fail, and faster decisions on whether major technology providers should fall under direct regulatory oversight. Firms will be expected to demonstrate how their AI behaves under market stress, data shocks, and feedback loops – and critically – to show who is accountable when automation operates at scale. More – not less – “human in the loop”.

 2. APAC: Taking the lead on Trustworthy Agentic AI

While the UK continues to debate the scope and pace of AI oversight, APAC regulators are moving decisively toward enforceable assurance regimes with direct implications for trading. Singapore has led the shift with the launch of its Model AI Governance Framework for Agentic AI, announced by IMDA at the World Economic Forum in January 2026, setting out expectations for accountability, bounded autonomy, lifecycle risk management, and testing of systems that can independently observe, decide, and act – read more here: https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/press-releases/2026/new-model-ai-governance-framework-for-agentic-ai. This is reinforced by practical LLM testing guidance aligned to emerging ISO/IEC AI management standards (https://www.iso.org/standard/42001) positioning assurance and explainability as operational requirements rather than policy.

This framework is not developing in isolation. Singapore is coordinating regional red-teaming and live testing across ASEAN, China, India, Japan, and Korea, and this week MAS and Korea’s AI Safety Institute formalised cooperation through a new MoU, publishing findings from joint testing of AI agents conducted over the past month (https://sgaisi.sg/resources/testing-ai-agents-for-data-leakage-risks-in-realistic-tasks/). In parallel, India’s January 2026 “techno-legal” AI white paper mandates that compliance be built into systems by design, including advance user notification, clear labelling of AI-generated content, formal risk assessment above compute thresholds, and regulatory investigation powers (India AI governance overview). South Korea’s new AI law similarly adopts a risk-based model emphasising trustworthy AI, transparency, standardisation, and strong supervisory oversight (Korea AI law analysis).

In APAC, firms that treat compliance as a post-hoc exercise look likely to find themselves out of step with supervisory expectations. AI compliance can no longer rely on governance frameworks and model documentation alone. Regulators are increasingly expecting controls to be embedded directly into trading systems: real-time observability, audit trails, stress and scenario testing, drift detection, access controls, and clear human override and accountability when automation operates at scale.

3. What This Means for Firms:

Most AI currently deployed in financial markets remains bounded by human intent: models forecast, optimise, and automate execution, but humans still authorise decisions. The new regulatory inflection point comes with the agency given to systems – those that independently decide, execute, and adapt in real time, from multimodal inputs and other systems without human intervention at the point of action. Once that threshold is crossed, markets begin to resemble adaptive machine ecosystems rather than human-led institutions augmented by software. While systematic trading is well established; what is changing is the grey line between fully autonomous decision-making and agentic large-scale interaction in a shortened feedback cycles.

Hedge funds have led adoption, benefiting from leaner governance, fewer legacy dependencies, and direct accountability for outcomes. These conditions allow AI to be embedded across research, execution, and portfolio construction, delivering alpha through ideas, new signals and scalability per employee. Banks are realising benefits as well, but progress is slower and more incremental, with most applications focused on productivity, analytics, and client service rather than autonomous decision-making given the risks involved. But this is creating structural differences across the industry: uneven deployment, control frameworks that cannot keep pace with system change, and a widening gap between technical teams and senior decision-makers internally as well as individual firms. Far from AI democratising, it is likely to create more of an unlevel playing field for market participants.

4. Lessons from Algo Trading

One way to address this imbalance is the work currently underway in the FIX AI and Algo Working Groups to extend previous insights from decades of algorithmic trading into emerging agentic workflows and lower cost of implementation. Familiar challenges – non-deterministic behaviour, timing variability, interaction effects, and regulatory interpretability – are resurfacing, but at far greater scale and speed. Sources of non-determinism now include system latency, complex interdependencies, feedback-driven behaviour, hybrid deterministic and probabilistic components, and continuously learning models. While these are classic non-linear dynamics, it’s the operation across more layers and under tighter time constraints that create new challenges. Traditional algorithms optimise execution within predefined strategies. Agentic systems extend autonomy upstream and downstream, coordinating specialised agents across research, portfolio construction, risk management, and optimisation. These systems adjust portfolios holistically and continuously, rather than reacting individually. The unresolved issues are operational: how to structure agent hierarchies, determine the appropriate number of agents, define tool access and prompting strategies, and evaluate performance at the system level rather than the individual component level – join the discussion and debate by emailing fix@fixtrading.org.

5. Reframing Data, Testing & Governance

If the reliability of agentic systems depends on the integrity of their data, feedback loops, and governance over time – data quality has to be maintained continuously across the full modelling lifecycle rather than treated as a one-off control, as even minor defects can quickly compound into material risk (see BlackRock’s recent analysis: https://arxiv.org/abs/2512.05559). Similar to traditional algorithmic adjustments but with greater impact, changes in one component of an agentic system can impact elsewhere, driving the need for governed quality layers that span data input, model behaviour, and downstream outputs. This requires more realistic testing approaches that assess resilience, constraint adherence, and unintended consequences by exposing systems to real production dynamics rather than abstract benchmarks. Governance frameworks will have to evolve beyond static approvals to reflect the adaptive nature of learning systems. While AI can replicate historical patterns, it does not recognise when objectives, constraints, or market regimes have shifted. Sustaining alignment in these conditions is increasingly an executive responsibility. Emerging practices such as continuous AI auditing and oversight of agent interactions provide clearer accountability, reduce systemic risk, and offer assurance that regulators and market participants can interpret. Evidence from firms such as Jane Street suggests that many weaknesses only emerge under full-system, deterministic stress testing rather than conventional pre-deployment validation (https://blog.janestreet.com/getting-from-tested-to-battle-tested/).

Anthropic CEO Dario Amodei has suggested that AI models could be capable of performing “most, perhaps all” software engineering tasks end-to-end within the next 6–12 months – read more here https://www.linkedin.com/posts/davidtimis_ai-softwareengineering-automation-ugcPost-7419813883890278400-2aaj/?utm_source=share&utm_medium=member_ios&rcm=ACoAAARqE5IBeeTUq6qzU4Eqccq_UO6-OMeR6Ao. If that timeline holds, the shift goes well beyond today’s notion of AI copilots with the broader implication is that the human-in-the-loop is shifting upstream. Software engineering may be the first example, but as AI systems become agentic – able to observe, decide, act, and learn without human intervention – traditional assurances based on static oversight are no longer adequate. When AI starts generating outputs at unprecedented scale – understanding, intent, and responsibility will still be human obligations. Whether in software or financial markets, agentic systems create value only when they are observable, governable, and accountable to people. Without that, delegation does not enhance assurance; it creates unnecessary risk.  Read more: https://bit.ly/45pUgXW. The Human in the Loop looks set to stay for now.

As always, thank you for reading – and a continued happy January.
Rebecca