The Bank of England is reconsidering whether existing financial rules can manage agents acting across payments, trading, cybersecurity and operations. Deputy Governor Sarah Breeden argues that requiring a person to approve every action is unlikely to remain practical as systems operate at machine speed.
The regulatory challenge is to preserve accountability without pretending continuous human intervention is possible. Guardrails must increasingly sit inside architecture and markets rather than rely on someone watching every decision.
Agentic finance creates correlated behaviour and cyber exposure
More than half of financial firms surveyed in a 2026 Cambridge report were already adopting agentic AI, although much activity remains in internal automation, engineering and knowledge work.
Agents differ from traditional rules because they pursue objectives, chain actions and adapt. Several systems trained on similar data may respond to the same signal in the same way, amplifying volatility or liquidity stress.
Cyber capability is an immediate concern. Defensive agents can identify vulnerabilities and isolate threats, while offensive models can scan and exploit at similar speed. Open models may trail the frontier by only months, limiting comfort from restricted access.
Shared clouds, software and payment networks create concentration. An attack or faulty update can affect several institutions together even when each firm’s local controls appear adequate.
Resilience must be designed for simultaneous disruption
Traditional recovery often assumes one firm or service fails in isolation. The Bank is considering scenarios in which several institutions lose critical systems at once.
Options include stronger failover, rapid reconstruction of compromised cores and arrangements allowing one bank to provide basic functions for another during an outage. Those mechanisms require legal, data and operational preparation before a crisis.
Stress tests should include model failure, poisoned data, cloud loss and autonomous defensive actions that worsen disruption. Recovery teams need clean records and manual or alternate routes for essential payments.
At firm level, agents require scoped authority, deterministic transaction limits and logs connecting objectives to actions. A model should not expand its own permission because circumstances changed.
Third-party dependencies must be visible across the system. Regulators can examine whether many firms rely on the same model, security tool or agent framework and therefore share an unrecognised failure mode.
Cross-border coordination is necessary because markets and cloud services do not stop at the Bank’s jurisdiction. Shared exercises can reveal conflicting recovery priorities before several authorities attempt different interventions during the same event.
Market safeguards may need to act before people can
Breeden raised circuit breakers and kill switches capable of slowing or stopping trading when faulty models contribute to disorder. Such controls already have precedents in markets, but agentic systems may require triggers based on correlated behaviour as well as price movement.
A kill switch needs an owner, tested authority and a safe restart process. Stopping activity can itself create risk if critical liquidity or payments disappear abruptly.
The Financial Stability Board has proposed practices covering governance, lifecycle risk, cyber, ICT and third parties. They are not binding, but emphasise clear roles for critical AI functions.
Regulators should avoid prescribing one model architecture. Outcome-based requirements for containment, evidence, recovery and accountability can remain useful as technology changes.
Financial firms must also explain when an agent materially influences a customer decision. System-wide resilience does not replace fairness and recourse at the individual level.
The Bank previously considered existing rules sufficient; new capabilities have exposed gaps. That willingness to revisit assumptions is important.
Human responsibility remains essential, but it must move upward—from approving every micro-action to designing boundaries, monitoring outcomes and intervening when the system crosses them. Financial stability will depend on controls that operate at agent speed while preserving a clear line back to institutions and people who remain accountable.
