Where last week’s story was sovereignty and kill-switches, this week the regulatory conversation moved from whether agentic finance is coming to what it looks like once it arrives and what this will mean for the industry. Meanwhile, on the buy-side, the same underlying idea (agents doing the legwork, humans keeping the judgment) is showing up not as a compliance question but as a design choice for how trading desks increasingly share knowledge. Here’s what I learnt on AI in Trading this week:

1. The FCA’s Mills Review: Sarah & Harry Meet #AskSheldon

The Mills Review -the FCA’s assessment of how AI will transform retail financial services by 2030, led by Sheldon Mills -was published on 6 July 2026. Team member William Penwarden shared his take on LinkedIn.

The core idea is “agentic finance,” illustrated through Sarah and Harry: a consumer whose AI assistant works overnight securing insurance ahead of a bike ride, flagging underperforming savings, and acting with permission rather than leaving her to do the legwork. The upshot is wider access to quality financial advice currently reserved for the wealthy – with two risks flagged specifically: market power and competition (concentration among model providers and hyperscalers, plus a coming “war for the consumer interface”), and financial crime and cyber risk (the same speed and scale that helps firms detect fraud also helps criminals commit it).

This landed alongside a wave of parallel signals from regulatory and government organisations this week right across the globe: HMT on strengthening the UK financial system, DSIT’s thematic review and gap analysis, the ESRB on frontier models and cyber risk, the Bank of England’s Financial Policy Committee, the UN’s Global Dialogue on AI Governance, and ASIC.

Why this matters for trading: Penwarden’s own focus in writing the review was the technology questions – how AI models, agentic systems, and automation could reshape the architecture of financial services itself: how decisions get made, how support is accessed, and how consumers, firms, and services connect around increasingly digital journeys impacts across all markets. Rather than write new rules, the FCA wants firms to apply existing duties to autonomous systems – but it also intends to extend its reach to firms currently operating outside the regulatory perimeter. A follow-up review will determine exactly how the FCA will treat general-purpose models and agent orchestration/gateway layers that sit outside current authorisation, with the Senior Managers Regime still binding a firm’s leadership even where a model’s behaviour and updates are partly outside its direct control.

The FCA is also building its own AI-enabled agentic supervisory model – applying to itself the same logic it’s asking of trading desks. @Nikhil Rathi linked this to an active pilot using agents to process market data at scale, floating a possible move toward 24/7 monitoring, and called for regulators to collaborate rather than seek new powers from Parliament.

2. The New Buy-Side: Norges’ Collective Intelligence Meets Millennium’s AI Lab and Amundi’s Fixed Income Push

Norges head quant trader @Peder Viervoll named a problem most desks recognise: the “archetype of the trader who keeps their alpha and their hard-won relationships in their head.” Rather than knowledge staying siloed, Norges traders now share domain knowledge with Claude, which encodes it into a skill file other traders can reuse – and an MCP lets non-technical traders query large databases once reserved for quants. The result, per Viervoll, is a “collective intelligence” where team members constantly challenge each other. Read more in Global Trading.

Millennium made the same bet institutionally, launching its own AI Lab to test and accelerate early-stage AI products with leading AI companies. And the Amundi Investment Institute (Mohamed Ben Slimane, Amina Cherief, Jiali Xu) published a comprehensive risk-return optimization framework for bond portfolios – a field far less developed than in equities – covering tracking-error decomposition, Markowitz optimization, decarbonization, and active share control, and flagging that ex-ante tracking error is generally overestimated for high-rated bonds and underestimated for low-rated ones.

Why this matters for trading: the “collective intelligence” framing is the buy-side’s answer to the same knowledge-silo problem the FCA is worried about at the consumer level – and fixed income trading is rapidly catching up to equity-style optimization, aided by AI tooling, all of which represents a significant structural shift in how PM and desk functions will interact in the future.

3. AI in Trading Is Now Mainstream – Ask Bloomberg

Bloomberg has launched a multi-asset Pre-Trade TCA API, giving clients programmatic access to pre-trade execution analytics directly within their own trading systems – embedding execution intelligence into workflows and scaling analytics across more order flow.

Why this matters for trading: this is the clearest signal yet that pre-trade analytics is moving from a terminal feature to embedded infrastructure – the kind of mainstreaming that tends to precede technology becoming table stakes rather than a differentiator.

4. It’s Not Just Secondary Markets – Primary Issuance Is Another Unsung AI Story

@Toby Nangle’s piece in FT Alphaville (Free to read )looks at primary issuance rather than the usual secondary-market story – prompted by the week SpaceX issued $86bn and the US$ investment-grade market printed $156bn without fanfare, followed by $129bn, then $200bn+. Nangle argues the corporate bond market is the real unheralded engine of primary capital markets, walking through a deal’s full lifecycle with input from Tareq Islam, Mark Lewellen (Deutsche Bank), Edward Farley (PGIM), Gustavo Baratta, Allan Malvar (ICMA), Rebecca Talbot (M&G Investments), and Nina Jaksic (PGIM).

Why this matters for trading: before any of the AI tooling in items 2 and 3 gets layered onto primary issuance, it’s worth understanding just how relationship-driven and manually intensive certain workflows still are today, which is exactly where future agentic workflows could help.

5. Governing the Agents – From Claude’s Hidden Workspace to Delegation Chains and Agent Protocols

Two threads converged this week around what it takes to trust an agent’s output and authority.

First, Anthropic reported evidence of an internal “workspace” inside Claude – dubbed “J-space” – where the model holds and moves concepts without writing them down, emerging unprompted during training. Researchers found real intermediate representations and showed they matter by altering one mid-thought and watching the output change. The pattern echoes global workspace theory, the leading account of how the brain broadcasts conscious thought – Stanislas Dehaene and Lionel Naccache, who built that theory, reportedly called the finding a landmark, though Anthropic makes no claims about consciousness. Conor Grennan’s viral analogy [link]: treating AI as software bolted onto existing workflows is why AI isn’t working – get data and interoperability right first, then treat the AI layer more like a merger than a software rollout.

Second, a piece on Evolving Agents introduced the Delegation Chain: the assumption that once a decision is approved the rest is “just execution” doesn’t hold, because every transfer of authority creates a fresh governance boundary requiring its own proof of authority.: the assumption that once a decision is approved the rest is “just execution” doesn’t hold, because every transfer of authority creates a fresh governance boundary requiring its own proof of authority. The infrastructure to support this is emerging: AGTP (Agent Transfer Protocol), a dedicated application-layer protocol with agent-native intent methods and protocol-level identity and authority; and Nomotic, covering agent identity, lifecycle, governance, context, orchestration, and intelligence. Alternatively, Palantir’s Governing AI Agents covers what agents can see and do, splitting governance into controls (authorization workflows, bounded execution, testing, evaluation, observability, fail-safe modes) and workflows (human-agentic collaboration and lifecycle development).

Why this matters for trading: this is the direct continuation of last week’s kill-switch data where 72% of banks couldn’t confirm they had one. Delegation chains, AGTP, and Nomotic and Palantir are early attempts to build the deterministic, always-available mechanism that gap requires -alternatively it could be industry led. The FIX AI Working Group is working on this and would welcome anyone who’d like to join the discussion.

Thanks for reading – as ever, any questions or feedback, let me know.

Rebecca