2026 appears set to be the year of Agentic Superpowers; specialized, customized AI trading agents are now outperforming general-purpose large language models in live crypto markets signalling significant progress over the last year (https://www.coindesk.com/business/2025/12/13/crypto-s-machine-learning-iphone-moment-comes-closer-as-ai-agents-trade-the-market).

However, the emerging end state is not fully autonomous, mass-market AI trading, but configurable AI portfolio managers where users define preferences and constraints while the system improves execution. Broad consumer adoption may be coming, but for now, as agentic AI matures from experimentation to deployment, three issues are becoming unavoidable: benchmarking that actually reflects value, interoperability between agents as a first-order systems problem, and the growing ambition for agents to design, configure and improve other agents. Alongside the perceived progress sits growing uncomfortable questions about infrastructure-level AI, market integrity and how regulation will be able to adapt when systems act at machine speed. This is what we learned this week on AI in Trading:

1. Building & Benchmarking Agents

Recent research from Recall Labs claims LLMs can trade autonomously when given the same prompts, but they only slightly outperform the market. The strongest results came from customized agents that added extra logic, better data sources, and tailored decision rules on top of base models in competitive trading environments (https://www.coindesk.com/business/2025/12/13/crypto-s-machine-learning-iphone-moment-comes-closer-as-ai-agents-trade-the-market). The research argues that the best agents did not focus only on profit and loss. They also used risk measures such as Sharpe ratio, drawdown, and VaR, aligning more closely with how trading teams already manage risk and performance. However, any claimed performance gain still depends on how it is measured.

Benchmarks only work when they reflect real business tasks, data, and risks. As @JamesCorcoran points out, only 16% of 445 AI benchmarks include uncertainty measures or statistical testing. Without these, many reported performance gains are unlikely to be meaningful.

As a result, firms are increasingly using internal benchmarks tailored to their own use cases. While this improves relevance, it makes comparison harder. Different models excel at different tasks, such as planning, coding, or reasoning, which is why teams are adopting multi-agent systems that route work to the most suitable model. Fine-tuning smaller language models is another option, but it is complex to operate and requires continuously updated real-world data.

A more practical alternative is to build agent-based systems that use tools and internal knowledge and are aligned with organizational rules and standards: what still matters is relevance to the problem at hand, not general performance scores. Read more here – https://lnkd.in/eBBf-TiX

2. From A2A to MAS ecosystems

When using Multi Agent Systems, how agents can communicate and coordinate from simple one-to-one interactions to swarm-based setups is creating an increasing interoperability issue. The launch of the Linux-backed Agentic AI Foundation, supported by OpenAI, Anthropic, Block and others, signals a clear aim to agree shared standards before the market fragments into incompatible proprietary systems (https://techcrunch.com/2025/12/09/openai-anthropic-and-block-join-new-linux-foundation-effort-to-standardize-the-ai-agent-era). Early work includes Anthropic’s Model Context Protocol, Block’s Goose framework and OpenAI’s AGENTS with AWS, Bloomberg, Cloudflare and Google also involved. The message is practical: no one wants to build custom integrations for every agent platform. Shared standards reduce integration effort for developers and allow enterprises to build agent systems that work across tools and vendors, rather than being locked into closed ecosystems.  We are starting to look at this in the FIX AI Working Group specifically for trading – if you’d like to get involved, please message direct.

3. The Agents building Agents Solution

As markets move from fixed algorithms to AI agents that can act autonomously and communicate with each other, a new challenge is emerging – building and configuring effective agents. Even strong models can perform poorly if they are misconfigured and fixing them often takes weeks of manual testing. TurinTech’s #Artemis addresses this by using AI agents to tune and improve other AI agents. If successful, this could reduce configuration friction while still aiming to preserve deterministic behaviour, which is essential in financial markets. Read more here: https://www.turintech.ai/devpreview

One reason generative approaches remain limited in trade execution is the need for determinism and immediate risk checks. At the CFM Imperial workshop this week (https://www.market-microstructure.net/program), Professor Justin Sirignano (https://www.maths.ox.ac.uk/people/justin.sirignano) highlighted his recent work on real-time recurrent learning (https://arxiv.org/abs/2501.08040), which could provide a foundation for continuous, online optimisation of recurrent models over long data sequences, addressing known limitations of truncated backpropagation. This could be particularly relevant for financial datasets such as limit order books, where sequences are long and model sizes are relatively small. However, improved learning fidelity does not by itself guarantee predictability or safety. That challenge remains one of governance. High-frequency trading works because behaviour is bounded, risk controls operate at machine speed, and failures shut down safely. If AI agents can be trained to build and tune other agents within these constraints, future real-time learning techniques could support greater autonomy without undermining the reliability that financial markets require.

4. Negative Implications When AI moves down the stack

The next shift may push AI further down the stack. Rather than supporting trading decisions, AI agents could operate at the infrastructure level, using reinforcement learning to act on early network-level signals. While this may resemble today’s HFT, the difference is material. Traditional HFT systems are human-designed to exploit known microstructure with fixed logic and bounded assumptions. Their behaviour is intentional, constrained, and largely predictable, even under stress. Infrastructure-level AI agents would not just execute strategies; they would discover them. Online reinforcement learning could enable far more frequent adaptation than human-designed approaches, optimising on low-level signals such as timing, sequencing, and transient queue states. These signals are mechanical rather than informational, but they could materially affect execution.

The real change is the move from explicitly engineered behaviour to machine-discovered strategies with limited human oversight. This heightens the risk of fragile strategies, feedback loops, and failures that only emerge in live markets, concerns highlighted as early as the UK Government’s Foresight Project, The Future of Computer Trading in Financial Markets (2012), and Andy Haldane’s paper in 2021 (https://www.bankofengland.co.uk/working-paper/2021/comparing-minds-and-machines-implications-for-financial-stability)

This debate raises questions around fairness, best execution, transparency, and systemic risk. The new regulatory direction is clear: supervisors will need stronger in-house AI capabilities, tighter cross-border coordination, and tools that can intervene at machine speed.

5. AI Regs and Moratoriums

Finally, I haven’t been able to keep up the separate AI for Planet newsletter over the past few weeks but in the meantime almost every week brings another article on the energy implications of AI, which is why this story stood out.

In the US, Trump’s widely heralded push to simplify AI regulation now sits alongside a very different response. A coalition of national, regional, state, and local organisations, representing millions of Americans, is urging Congress to impose a national moratorium on approving and building new data centres. The group argues that the rapid expansion of data centres, driven largely by generative AI and cryptocurrency, poses significant environmental, economic, and social risks with electricity demand could rival that of 30 million households within five years, alongside water use for cooling equivalent to 18.5 million households. The coalition argues that data centre growth remains largely unregulated and is already disrupting local communities while undermining climate, water, and energy security. Until comprehensive safeguards are in place, the signatories are calling for a halt to further development. The appeal is led by Food & Water Watch and backed by a broad alliance of environmental, climate, public health, consumer protection, labour-aligned, and community organisations across all 50 states. One to watch, to see which side ultimately prevails. Read more here: https://lnkd.in/eQfqPrKG

As AI trading tools scale, any advantage is likely to be short-lived and driven by whoever can innovate fastest, repeating earlier waves of trading tech, but with a wider gap opening between leaders and followers. AI in trading is now splitting in two directions: economic AI that improves risk management and price discovery, and infrastructure-level AI that optimises latency, inference, and queue position, often benefiting individual firms while potentially weakening market integrity. How autonomous these agentic systems become, and how deeply they are embedded in market infrastructure, will shape how supervision and governance evolve. Without clear constraints, shared standards, and oversight that operates at machine speed, markets risk handing control to systems optimising the wrong objectives at scale. One to watch for Market Microstructure in 2026.

Thank you for reading. I’d love to hear your thoughts – what you agree with, where you don’t and what you’d like to see more of next time.

Best wishes,
Rebecca