en.Wedoany.com Reported - LaunchDarkly officially launched AgentControl on May 19, a new solution designed specifically for providing real-time management capabilities for AI agents in production environments. As more AI agents are deployed into production, enterprises are discovering that generative AI is probabilistic in nature, and its behavior does not always perfectly match the testing phase—even if the code remains unchanged, models can still experience issues such as declining output quality or deviation from assigned tasks when facing complex user interactions. AgentControl was created precisely to address this pain point.

The core design philosophy of AgentControl deeply integrates real-time runtime intervention with the complete operational layer required for the reliable functioning of AI agents. The platform supports configuring agent behavior across teams and frameworks, enabling quality benchmark assessments before changes impact live traffic, and achieving controlled deployments through progressive rollouts and guarded launches, all without the need to redeploy the underlying application. This means that whether they are IT teams, software engineers, AI engineers, or business users building AI agents with no-code tools, everyone can maintain and manage agents on the same platform, significantly lowering the technical barrier to agent operations.

Real-time response speed is a key competitive advantage of AgentControl. Platform configuration changes can be propagated within 200 milliseconds, a speed sufficient to adjust agent behavior, switch to different models, or trigger fallback strategies in real-time during an ongoing conversation—the entire process is completed before the user perceives any anomaly. LaunchDarkly CTO Cameron Etezadi commented on this: "The thorniest problems in the AI space, such as model drift, output instability, and insufficiently timely intervention responses, are precisely what our platform excels at solving. We didn't need to rebuild the platform; we just needed to extend it to accommodate the new demands of the AI software development lifecycle and agent-driven workflows."

AgentControl provides enterprises with a complete control chain from configuration and evaluation to launch. Teams can use experimentation features to compare the performance differences of different model versions on metrics like cost, latency, and user satisfaction, automatically score output quality in completion mode through online evaluation mechanisms, or programmatically invoke judges in agent mode via the AI SDK. The platform also supports setting differentiated safety filtering rules for different user types, geographic regions, or application scenarios, and continuously tracks model behavior changes through metric monitoring. These capabilities enable enterprises to safely evaluate model behavior in production, run targeted experiments, and gradually introduce new models, avoiding lock-in with a single model vendor.

Brian McCarthy, President of Global Revenue and Field Operations at Anysphere Inc., the developer of the Cursor AI code editor, pointed out: "As more AI-driven products and agent capabilities enter the production phase, runtime control has become as critical an infrastructure component as the development workflows and governance mechanisms teams already trust."

Founded in 2014 and headquartered in Oakland, California, LaunchDarkly was co-founded by Edith Harbaugh and John Kodumal and has long focused on providing feature management platforms for software development teams. The release of AgentControl marks the company's comprehensive extension from traditional feature control into runtime governance for the AI era. As development teams continue to deploy more agents that run, write, and interact at machine speed, the corresponding control systems must also possess the rapid response capabilities to match.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com