LaunchDarkly launches AgentControl for AI agent operations

LaunchDarkly has launched AgentControl for software teams managing AI agents in production, extending its platform into AI operations.

AgentControl is designed to let teams change how an AI agent behaves without redeploying the underlying application. The system can alter an agent's runtime behaviour in less than 200 milliseconds, including switching to another model or triggering a fallback response before a user sees an unsatisfactory answer.

The launch reflects a broader shift in software development as businesses move AI agents from experimentation into live, customer-facing services. That transition has created operational issues that differ from those in conventional software releases, because agent behaviour can change according to prompts, models and production context even when the core application code remains the same.

Once deployed, agents' settings and management processes often sit across multiple teams and frameworks. That fragmentation has increased demand for shared rules on governance, version control and release management, particularly where engineering teams need to intervene quickly as live conditions change.

AgentControl combines runtime intervention with a broader operational layer for teams running AI agents. It offers central oversight across teams and systems, benchmarking before changes reach customers, progressive roll-outs, monitoring with trace-level visibility, and the ability to adjust agents based on production data.

LaunchDarkly built its name in feature management and software release controls, giving engineering teams a way to alter software behaviour in production without a full redeployment. With AgentControl, it is applying that approach to AI systems, where intervention may be needed during a conversation rather than over a conventional deployment cycle.

That matters because AI agents can produce inconsistent outputs, while model performance may shift over time. Businesses using these systems have been looking for ways to monitor such changes, test updates more carefully and reduce the risk of errors reaching end users.

In practice, the product is intended to cover both direct control and broader oversight. Teams can change configuration at runtime, monitor agent performance live and set standards for how agentic workflows are governed across different parts of an organisation.

Cameron Etezadi, Chief Technology Officer at LaunchDarkly, said the company saw clear overlap between established software control problems and the operational challenges now emerging around AI agents.

"LaunchDarkly has always been about giving software teams control at runtime over what their software does in production," said Etezadi.

"The hardest problems in AI, like model drift, unpredictable outputs, and the inability to intervene fast enough, turn out to be exactly the problems our platform was built to solve. We didn't have to reinvent the platform; we just had to extend it to meet the demands of an AI SDLC and agentic-driven workflows," added Etezadi.

The company also pointed to support from Cursor, which develops AI coding tools used by software teams. Both businesses are positioning runtime controls as an increasingly important part of AI software delivery as more organisations move from pilot projects to production deployments.

"Cursor is how the world's leading enterprises are building with AI. As more AI-powered products and agentic capabilities reach production, runtime control becomes essential infrastructure alongside the development workflows and controls teams already trust," said Brian McCarthy, President of Global Revenue and Field Operations at Cursor.

"LaunchDarkly built an additional layer for that environment, and AgentControl extends it to the agent lifecycle in a way that complements how Cursor's customers already build," added McCarthy.

LaunchDarkly says it serves thousands of customers worldwide, including a quarter of the Fortune 500, from its base in Oakland.