LaunchDarkly deploys agent fleet and launches runtime control product

LaunchDarkly, a feature flagging company with 600 employees, is using AI agents to transform its software development lifecycle. The company has deployed a fleet of over 20 agents to work on specific codebase refactoring tasks, shifting engineers' roles toward reviewing pull requests and acting as architects. This approach requires decomposing work into atomic units with explicit success criteria, as agents lack inherent concepts of 'good' code. LaunchDarkly also utilizes a long-running agent named Vega for automated flag cleanup, identifying and deleting stale flags.

The company cites Anthropic's recent blog post noting a shift from 2,500 to 30,000 daily changes as evidence of the need for autonomous rollback mechanisms. LaunchDarkly argues that feature flags enable automated rollbacks within seconds, allowing for high-velocity deployment without human babysitting. The company emphasizes that while agents can handle implementation and bug fixing, human judgment remains necessary for compliance and strategic product decisions.

LaunchDarkly has introduced a product called Agent Control, which provides runtime guardrails for AI agents. This tool allows developers to switch models or prompts without redeploying software, enabling dynamic adjustment of agent behavior based on real-time telemetry. Key features include the ability to shut down or dial back agents if model drift occurs, error rates spike, or token consumption becomes excessive. The system supports A/B testing of different prompts or models across user segments. Edith Harbaugh positions Agent Control as a critical infrastructure layer for ensuring the safety, cost-efficiency, and reliability of AI agents in production, analogous to feature flags for traditional software.