Cameron Etezadi, CTO, LaunchDarkly
“It’s hard to imagine a more timely or important issue,” Bank of England Deputy Governor Sarah Breeden noted recently on the rise of agentic AI.
On the speed of change, I couldn’t agree more. Over the past few years, autonomous agents have reshaped cyber risk, trading, and payments faster than regulatory frameworks can keep up. Just a few years ago in 2024, software task completion capabilities doubled every four months driving a rapid leap toward autonomous operations. Now they’re many times faster. Today, financial institutions are deploying AI across core functions with a vast majority giving these tools higher degrees of autonomy.
As exciting as this is for the industry, the rapid adoption of AI has resulted in new threats on several fronts. Malicious actors have used agentic AI to exploit cyber vulnerabilities at speed, escaping sandboxes to hack systems and reverse-engineering software patches faster than firms can deploy them. Meanwhile, rising debt-financing across AI infrastructure has increased the potential shock to financial stability if asset prices drop.
However, the proposed solutions offered in Sarah Breeden’s speech – such as market-wide ‘kill switches’ – do not align with the reality of modern financial infrastructure. In an industry built for always-on availability, pulling the plug is a self-inflicted outage. Rather than asking how to stop AI, the conversation must explore how to safely monitor, adjust, and govern autonomous agents in live production.
This necessity is because governance alone isn’t enough once agents are live. Current enterprise AI discussions focus heavily on models, data quality, and static compliance frameworks, but these address what AI should do rather than what it is doing in production. Traditional authentication models were built for human confirmation, not for software delegated to transact independently. As unforeseen scenarios emerge, organisations need the ability to intervene and adapt behaviour in real time.
The principles of stability
Maintaining stability across the financial sector requires three core operational principles: guarded rollouts, real-time monitoring, and runtime controls. These capabilities offer a resilient alternative to emergency shutdowns.
Just as feature flags transformed software delivery by letting developers modify application behaviour without redeploying code, AI systems require equivalent runtime controls. Modern financial infrastructure depends on granular operational levers over live agents. Financial institutions must first execute guarded rollouts, testing autonomous agents on low-risk tasks before expanding access to complex trading or payment workflows.
Continuous real-time monitoring then allows teams to dynamically alter prompts, swap underlying models, or re-route requests the moment an agent deviates from its intended path. When market conditions turn volatile or autonomous agents begin herding toward dangerous execution patterns, real-time observability gives operators instant awareness of anomalous activity.
Finally, runtime controls enable instant fallback behaviours, such as reverting an agent to read-only status or routing transactions back to human oversight. Better yet, this happens without taking core banking systems offline. This granular intervention prevents catastrophic, simultaneous multi-firm outages. As seen during major software disruptions, emergency patches themselves can trigger widespread operational failure if deployed recklessly. Runtime controls allow security teams to safely isolate compromised agent behaviours or apply patches incrementally, maintaining continuous operations without taking core banking systems offline. In almost all cases, a slight change or degradation in functionality is a far superior outcome to an outage.
Establishing these three principles allows financial institutions to preserve sector stability while maintaining rapid innovation. Precise operational mechanisms deliver the control required to deploy capable agents while safeguarding systemic resilience, compliance, and customer trust.
Balancing autonomy with oversight
To put these principles into practice, financial institutions must match increased agent autonomy with immediate human intervention. When live model drift occurs, waiting for traditional engineering cycles, governance approvals, or infrastructure redeployments exposes firms to unnecessary risk. Teams require the ability to modify live application behaviour instantaneously.
Practical execution requires establishing continuous operational oversight across every deployed system. Centralising visibility across all active agents provides the clarity needed to monitor real-time performance and enforce compliance consistently across disparate teams. Combining this visibility with runtime controls gives financial leaders the operational framework to satisfy stringent regulatory expectations while maintaining accountability and system reliability.
Leaders at the Bank of England and beyond are right to highlight the undeniable truth that agentic AI is moving fast, and regulators are struggling to keep up, but threatening to pull the plug with market-wide kill switches simply won’t work in a 24/7 financial ecosystem. Thinking otherwise misestimates the cataclysmic risks of such a shutdown to the economy. Instead, the correct answer is smarter controls, not blunt instruments. Institutions that have the power to observe, adjust, and govern autonomous agents in real time will have unshakeable stability whenever the next technological leap arrives.