Questions continue on who holds the controls when an agent acts – and can you prove it
Last week’s question was how far AI is allowed to act. This week Jon Hall, speaking independently, returned to the agents he first warned about in his 2024 speech Monsters in the deep?. In Principals and Agents he framed them as a principal–agent problem: delegation saves time, but the firm, not the agent, carries the residual cost. His conclusion was deliberately unglamorous: no new rules – just apply the existing ones properly, to every firm that delegates to an agent.
The week’s news:
• State Street linked LINK to Morgan Stanley Matrix, putting a bank’s in-flight TCA and algo controls inside a custodian’s FX platform.
• Coinbase and Schwab opened accounts to agents with limits and logs held by the venue and the broker, not by whoever built the agent.
• Credit: Man Group, AllianceBernstein and LTX told FundFire how AI is moving deeper into bond research and trading workflows – with LTX stressing that asset managers remain accountable for actions taken on their behalf.
• Swarm risk: Hall warned that agents built on the same few models could fail, de-risk or be switched off at the same moment, draining liquidity just when it is needed.
• Gravitee found only 6% of firms can revoke a rogue agent’s access within an hour – the basic test of whether an agent kill switch works.
• AIMA reported that only 7% of fund managers call generative AI core to portfolio construction currently but AI is gradually shaping orders before they reach the desk.
Most of the controls in this week’s news sit with a platform – a broker, a bank’s analytics, a vendor’s agent gateway – not with the firm whose name is on the order. RTS 6 was built for the firm sending orders to the market using its own logic; increasingly, that logic arrives from another platform. Evidencing your controls may mean evidencing someone else’s. RTS 6 remains a strong control framework – the gap opens where decision-making, tools and permissions extend beyond the order execution layer.
Here are the five themes I learnt this week on AI in Trading:
1. Two years on: same rules, applied properly – and across the board
In May 2024, Jon Hall’s Monsters in the deep? asked whether neural networks could become “deep trading agents” – and warned that, if they did, markets could become more brittle and correlated, and agents’ incentives could drift away from those of regulators and the public good. His follow-up, Principals and Agents moves the theory on in two ways: the agents are no longer hypothetical, and the focus shifts from training the model to controlling how firms delegate to it – treating an agent as “an AI model wrapped in a harness”, with identity, permissions and monitoring around it. So what’s different two years on:
· What has surfaced. Autonomous trading by AI agents has begun – led by retail investors in non-core markets such as crypto, and enabled by MCP connections – and agents have shown collusive behaviour in market pricing experiments. He still describes this as predominantly a discussion of future risks.
• What has not changed. In 2024 he wanted agents to comply with both the spirit and the letter of regulation, and warned that the trading manager would be held responsible. In 2026 he is explicit that he is not arguing for new regulations: agents should be trained to respect “the very same rules and regulations as a human employee”, and firms remain responsible for their agents’ actions.
• What has moved on. 2024 focused on the model: train it, test it against other agents, stress test it adversarially. 2026 adds the harness around it: a unique identity building on the algo IDs in UK RTS 22 Article 8(3), a least-privilege sandbox, permissions “neither overly broad nor overly persistent”, and scalable, risk-based monitoring.
• From the firm to the system. He now adds market-level monitoring: concentration among service providers, correlation through market intelligence, and multi-agent simulations such as the BIS’s Project Logos, launched in June, which compares LLM-based and rules-based portfolio-manager agents.
Applied properly, the existing rules already reach further than many assume. Under RTS 6 Article 4, a firm remains fully responsible where it outsources or procures software or hardware used in algorithmic trading. In the EU, ESMA expects firms to keep the capacity to monitor, suspend or terminate algo trading without depending on the provider. The catch is that RTS 6 controls act on the orders a firm sends to market – not on decisions taken upstream on someone else’s platform.
Why this matters for trading:
• No new rules is not a reprieve. Best execution, market abuse and RTS 6 already apply wherever an agent’s activity falls within them – there is no new regime to wait for. Hall places responsibility on the firm and does not address vendors directly. But RTS 6 already says outsourcing does not move responsibility – which brings the harness an agent runs in, including those built into EMS, TCA and broker platforms, into the same frame as an algo.
• The algo ID becomes the agent ID. RTS 22 already identifies the algorithm behind an order; Hall’s proposal extends the same logic to agents, together with the ability to revoke their access.
2. The controls now sit on someone else’s platform
This week’s launches put the limits, logs and switches in the hands of platforms rather than the firm placing the order:
• State Street and Morgan Stanley: State Street integrated LINK with Morgan Stanley Matrix, giving eligible buy-side FX clients the bank’s in-flight TCA and algo controls for trades executed via FX Connect, with single sign-on.
• Bloomberg: its sell-side EMS is becoming a central control plane where banks plug in their own pricing models, algos and AI.
• Coinbase: Agent Controls cap each trade, limit 24-hour volume and stop the agent at a set loss – enforced on Coinbase’s servers, not by the agent.
• Charles Schwab: its MCP pilot authenticates the agent, sets its scope and logs every action – while the agent itself runs on an AI platform Schwab does not own.
Credit is seeing the same shift, with AI moving deeper into research and execution workflows on platforms the asset manager does not run:
• Man Group: AI tools – in particular Claude – are now “integral” to its credit research, analysis and portfolio management, Sriram Reddy told FundFire. The firm has used AI to reclassify more than 6,000 companies by sector and identify the BDCs most exposed to AI disruption.
• AllianceBernstein: Timothy Kurpis, head of investment-grade credit, said LLMs now surface the key risk factors in idiosyncratic data-centre deal documents that used to take days to review – while not telling analysts how to judge an investment.
• LTX: CEO Jim Kwiatkowski said asset managers need to know how their data is secured, how AI answers are verified and how actions taken on their behalf are performed: “There has to be an audit trail because asset firms are still accountable for the end result.”
Why this matters for trading:
• The broker controls the order, not the decision. Pre-trade controls fire at order entry. A price collar can stop a bad order from an agent on a vendor’s platform – but it cannot see why the order was sent.
• Evidence is split across platforms. Best execution sits with the asset manager, the logs with the platform, the kill switch with the broker. No single firm holds the full picture of who controls what – or who can pull the switch.
• Dependence on the provider. ESMA’s expectation that firms can suspend algo trading without depending on the provider is harder to meet when the agent, its limits and its logs all sit on someone else’s platform.
• Accountability travels with the bond trade. LTX’s point – the asset manager stays accountable for actions an AI tool takes on its behalf – is the same split between control and responsibility, now reaching credit workflows where data, pricing and RFQs already sit on third-party platforms.
3. Correlated by design: the swarm problem
For heads of dealing, the most important part of the speech is what happens when many firms delegate at once. Hall treats liquidity as a common good provided by a swarm:
• Three dynamics weaken it: concentration among providers, correlation between them, and correlation that rises under stress.
• Shared foundations, shared failure. Agents on similar models, common failures and managers switching agents off after surprising behaviour could all hit at the same moment.
• Collusion without communication. Access controls can block direct channels, but not tacit collusion or momentum ignition through prices and timing.
Hall’s answer is to measure it: track how concentrated and correlated AI service providers are. A separate BoE–FCA survey shows why. The top three model providers account for 54% of firms’ model relationships, and the top three cloud providers for 79%. Firms have deep knowledge of only 24% of the third-party AI systems they use, relying on vendor assurance for 81%. And interconnectedness – automated trading, common failure modes – is the systemic risk firms rate lowest today but expect to grow fastest.
Why this matters for trading:
• Kill switches can become liquidity demand. If everyone switches off together, the switch-off itself drains liquidity. Single-firm stress tests are not designed to capture that scenario.
• Model concentration is a best-execution issue. If a firm’s algos, its brokers’ and its TCA all run on the same few models, independence in best-execution evidence may not hold under stress.
• Agents may find manipulation on their own. Momentum ignition is a behaviour a profit-seeking agent could discover on its own, without any human intent behind it.
4. Can you switch it off within the hour?
A kill switch is only a control if it works, fast, and leaves a record:
• Gravitee: a survey of 900 US and UK technology leaders found only 6% can revoke a rogue agent’s access within an hour, and only 1 in 8 can fully reconstruct what an agent did 30 days ago. 58% have already revoked an agent during a live incident; only 25% then held a named person responsible.
• Financial services is not immune: it reported the highest incidence of agent-related security incidents of any sector, at 66%, despite comparatively strong controls.
Why this matters for trading:
• Time-to-revoke is becoming a metric. Gravitee’s figures suggest few firms can cut one agent off quickly across every system it touches – OMS, EMS, broker portals, data – particularly those run by third parties.
• Reconstruction is the audit test. Without a replay of what an agent did last month, best execution for the orders it shaped is hard to evidence.
• Limits outside the agent’s reach. Coinbase enforces limits on its own servers, so the agent cannot change them – the same separation of duties that already applies to traders and algos.
5. AI reaches the PM before it reaches the order
Two surveys show where buy-side AI actually sits today – and it is upstream of the desk:
• AIMA: a survey of more than 100 fund managers with around $998bn in AUM found generative AI use falls to 31% in portfolio construction, though only 7% consider it core. 60% say explainability matters for AI-assisted decisions.
• BoE–FCA: even including plans for the next 12 months, no more than 9% of firms expect agents to execute trades or place orders.
• Citadel: at Bloomberg’s Enterprise Data & Tech Summit, Andrew Janian described giving teams room to experiment while preserving human judgement in the most consequential parts of the investment process.
Why this matters for trading:
• The order may already be AI-shaped. If research and idea generation run through AI, the desk receives orders whose rationale it cannot see. The desk’s challenge on urgency and sizing becomes one of the few human checks on that rationale.
• The trader is the last human check. With agents kept away from execution, the dealing desk is where AI-informed decisions first meet the market.
• Explainability will flow downstream. If PMs need to explain AI-assisted decisions, the question of how those orders were worked is likely to follow them to the desk.
None of this week’s news needed a new rule. It needed an owner, a log and an off switch for every agent – and the ability to prove all three when the platform that holds them is not yours.
Thanks for reading as always.
Rebecca


