LaunchDarkly launches runtime control layer for the agentic AI era

LaunchDarkly, a feature control platform that helps developers and software engineers launch and manage products, today announced the launch of AgentControl, a new solution providing real-time management of AI agents in production. 

As more agents get launched into production, companies are discovering that, unlike ordinary code, generative AI is probabilistic and doesn’t always behave exactly as it did in testing.  

Although models can update and go through experiments, they might run into unexpected user interactions that change output quality without a single change to code. Once deployed, agents are given broad autonomous leeway, with less human oversight, primarily so they can get their job done. But that leaves them open to straying from their tasks as anomalies build up, and it can happen faster than teams can react.  

“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,” Chief Technology Officer Cameron Etezadi said. “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.” 

LaunchDarkly said its platform is built to combine instantaneous runtime intervention with an operational layer that teams need to run AI agents reliably. It can assist with configuring agentic behavior across teams and frameworks, benchmark quality before changes reach live traffic, and provide progressive releases for guarded rollouts.  

All of this can be done with closely monitored trace-level visibility, allowing teams to iterate based on what production data actually shows. That means information technology teams, software engineers and AI engineers, alongside business users building AI agents using no-code tools, can slow-roll agent maintenance all without the need for redeployment. 

LaunchDarkly says the kind of control needed to manage agents in production requires a certain kind of speed. Configuration changes propagate in under 200 milliseconds, fast enough to adjust the behavior of an agent, route to a different model, or trigger a fallback – all within the turn of a conversation, and before a customer even notices something went wrong. 

“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 Anysphere Inc., maker of the Cursor AI code editor. 

LaunchDarkly has a long history of building powerful control and management interfaces for software engineering teams. Its largest audience is development and operations, or DevOps, where adjusting the feature configuration for applications is part of the company’s journey.  

AgentControl represents that earned expertise applied to the new era of agentic AI, providing companies with a powerful way to keep agents in check. As development teams continue to deploy more agents that write, deploy and interact at machine speed, the company argued that they also need a system that can react with the same alacrity. 

Image: SiliconANGLE/Microsoft Designer