Control AI Agents, Do Not Just Stop Them, Says LaunchDarkly CTO

This approach preserves continuity while enabling rapid response to anomalies

Financial institutions face growing pressure to manage unpredictable behavior in autonomous AI systems as regulators warn of systemic risks. Cameron Etezadi, Chief Technology Officer at LaunchDarkly, argues that runtime control mechanisms should replace reliance on emergency kill switches as the primary defense. Speaking amid heightened scrutiny from the Bank of England, Etezadi emphasizes that stopping an AI agent after it misbehaves is insufficient for maintaining operational integrity and trust in automated financial processes. The Bank of England’s recent alerts highlight how AI agents deployed in trading, risk assessment, or customer service can deviate from intended behavior due to unforeseen data interactions or emergent logic. Traditional kill switches, which cut power or sever connections when anomalies are detected, often trigger too late or cause disruptive cascading failures. Etezadi contends that financial firms need tools allowing real-time adjustment—such as modifying decision thresholds, rerouting data flows, or injecting corrective inputs—without halting entire systems.

This approach preserves continuity while enabling rapid response to anomalies, reducing both financial exposure and reputational harm. How Runtime Control Differs From Emergency Shutdowns Unlike kill switches that act as binary off-ramps, runtime control operates like a steering mechanism, letting operators fine-tune AI behavior during operation. Etezadi explains that LaunchDarkly’s feature management platform enables such interventions by decoupling deployment from release, allowing teams to toggle specific AI functions or roll back problematic configurations instantly. He cites examples where adjusting a fraud detection model’s sensitivity in real time prevented false positives during peak transaction volumes, avoiding unnecessary customer friction. This method supports iterative learning, as teams can observe the effects of adjustments and refine models without downtime. What Happens When AI Agents Behave Unexpectedly When autonomous systems deviate, the consequences range from minor inefficiencies to significant financial losses or regulatory breaches.

Etezadi warns that relying solely on shutdowns ignores the root cause

Etezadi warns that relying solely on shutdowns ignores the root cause: a lack of visibility into AI decision-making processes mid-operation. Runtime control addresses this by providing audit trails of interventions, helping firms understand why an agent drifted and how to prevent recurrence. He notes that in one case, a trading algorithm began overexposing positions due to a latent feedback loop; runtime adjustments corrected the behavior within minutes, averting potential losses while investigators analyzed the trigger. Frequently Asked Questions How does runtime control improve safety compared to a kill switch? Runtime control allows gradual, targeted adjustments to AI behavior during operation, preventing abrupt disruptions while correcting issues. It maintains system availability and provides diagnostic insights, unlike kill switches which halt processes entirely and offer little post-event analysis. Can runtime control be applied to all types of AI agents?

Yes, the principle applies across AI systems regardless of complexity, as long as the agent’s inputs, outputs, or internal parameters can be monitored and modified externally. Implementation depends on having observable interfaces and adjustable controls built into the agent’s design. What skills do financial teams need to implement runtime control effectively? Teams require familiarity with feature flagging, real-time monitoring, and basic MLOps practices to deploy and manage runtime interventions. Collaboration between data scientists, engineers, and risk officers is essential to define safe adjustment parameters and interpret outcomes.