Sarah Breeden delivered a stark message in Portugal this week. The Bank of England’s deputy governor for financial stability told an audience at the European Central Bank Forum on central banking that current oversight structures fall short when machines start deciding for themselves. Her words landed with force. They reflect growing unease among supervisors who once viewed artificial intelligence as just another modeling tool.
Agentic systems mark a break from earlier forms of automation. These AI setups don’t simply generate text or analyze data on command. They pursue goals, adapt strategies, interact with other programs, and execute actions with minimal human input. In financial markets that speed matters. A single trading agent could scan opportunities across venues, adjust positions in milliseconds, and coordinate with peers in ways humans cannot track. Breeden put it plainly. “Our frameworks were not built to contemplate autonomous agents, and relying on a human in the loop for all agent actions is unlikely to be realistic,” she said. “More sophisticated governance and accountability frameworks may be needed.”
The remark comes at a moment when banks and asset managers eye agentic tools for payments processing, compliance checks, credit decisions, and market making. Yet adoption remains early. The Bank’s Financial Policy Committee noted in April that firms have not yet deployed advanced forms in ways that create systemic danger. Risks, however, could climb fast. That April 1 letter to parliamentarians made the point clear. Evidence shows limited current exposure, but intent to expand use is rising. Further targeted study on agentic applications in payments and trading infrastructure is now under way.
Breeden’s concern echoes positions she laid out in a November 2024 speech hosted by the Bank for International Settlements. There she described generative models as capable of learning and evolving autonomously at speed. Outputs often resist easy explanation. Objectives may drift from those set by developers or society. The question she posed then still hangs over policy makers. As models gain power and spread into wider uses, can regulators keep depending on technology-agnostic rules written for earlier eras? Those rules assume human managers retain clear sight of what systems do. Agentic setups test that assumption directly.
So far the UK approach has emphasized flexibility. The Prudential Regulation Authority expects firms to meet standards on data quality, model risk, operational resilience, and third-party dependencies regardless of the underlying technology. Senior managers must still demonstrate they understand and can control outcomes. But Breeden and colleagues now openly review whether those expectations hold when decision loops run without constant human sign-off. The Bank and FCA established an AI Consortium in 2025 precisely to gather private-sector insight on these questions. The group examines concentration risks from shared models, edge cases that emerge in live deployment, transparency shortfalls in generative outputs, potential for rapid contagion across trading desks, and the specific behaviors of agentic systems. A report summarizing its findings is due later this year.
Market participants sense the shift. Global standard setters have issued parallel cautions. The Financial Stability Board called in June for tighter controls around AI agents, citing distinct challenges to traditional oversight. Analysts point to early demonstrations such as Anthropic’s Mythos as illustrations of fresh cybersecurity vulnerabilities that could affect banking infrastructure. One worry centers on herding. If multiple institutions rely on similar foundational models or agent playbooks, correlated behavior during stress could magnify sell-offs or liquidity drains. Stress tests that once focused on balance-sheet shocks may need redesign to simulate interactions among autonomous agents whose reaction functions remain partly opaque.
But regulators face a difficult trade-off. Overly prescriptive rules could slow beneficial innovation. Banks already use simpler AI to cut costs in routine tasks. Agentic versions promise bigger gains in areas like fraud detection or personalized advice. Premature restrictions risk pushing activity offshore or into less supervised corners of finance. The Treasury Committee has pressed the Bank and FCA to move beyond a wait-and-see posture. Lawmakers urged AI-specific stress tests and clearer expectations around accountability when systems act independently.
The Bank’s April response rejected any suggestion of complacency. Officials described their stance as proactive yet proportionate. They continue biennial surveys of AI uptake across supervised firms. Roundtables with chief risk officers and technology leads surface practical constraints. International coordination through the Financial Stability Board, Basel Committee, and G7 forums helps align approaches. Still, Breeden’s latest comments signal that patience has limits. Governance upgrades appear inevitable. Firms may need explicit policies on when agents must escalate decisions, how to maintain audit trails for autonomous actions, and mechanisms to pause or unwind harmful sequences.
Some ideas already circulate. Senior-managers regimes could add specific responsibilities tied to agentic deployments. Recovery plans might incorporate “kill switches” or circuit breakers calibrated to AI-driven events. Third-party model providers could face direct oversight if their offerings become critical infrastructure. None of these steps are settled policy. Each requires careful calibration to avoid unintended consequences. Yet the direction is visible. Technology-agnostic rules worked during the experimental phase. As agentic systems move from pilots to production at scale, that agnosticism will be tested.
Financial institutions watch closely. Those building internal agents report challenges around explainability, bias amplification, and unintended goal pursuit. Compliance teams struggle to map every possible output when systems can chain tools, query external databases, and rewrite their own subroutines. The gap between regulatory expectation and technical reality widens with each capability jump. Breeden’s speech at the ECB forum served as both warning and invitation. Supervisors stand ready to adapt, but they expect industry to surface problems early and share mitigation strategies through forums like the Consortium.
The stakes extend beyond individual firms. Financial stability rests on confidence that markets function predictably even under strain. If agentic trading amplifies volatility or payments agents create unexpected bottlenecks, that confidence erodes quickly. Cyber resilience adds another layer. Autonomous agents could be weaponized or compromised at machine speed, outpacing human response teams. Breeden highlighted these systemic angles in her earlier BIS remarks and returned to them this week. The message has consistency. Monitor, test, and prepare to adjust the rulebook.
Progress will not be linear. Technical advances arrive in bursts. Regulatory reform moves more deliberately, shaped by consultation and legislative timetables. The coming months will bring the Consortium’s report, results from expanded supervisory surveys, and likely further speeches that refine the Bank’s thinking. Firms that treat these signals as prompts for serious internal work stand to gain. Those hoping the issue resolves itself without hard choices face rising exposure.
Breeden did not call for immediate overhaul. She indicated instead that the conversation has entered a new phase. Frameworks built for static models cannot simply stretch to cover dynamic, goal-seeking agents. Governance must evolve in tandem. Accountability chains need clarity when no single person presses the button. And supervisors must retain tools powerful enough to contain spillovers if things go wrong. The speech in Portugal was measured. Its implications are not.
WebProNews is an iEntry Publication