June 2026 Summaries

Warehouse-native experimentation has been expanded to support BigQuery, Databricks, and Redshift, alongside the existing Snowflake integration, allowing organizations to run experiments directly on their trusted data without duplicating it or creating alternative data versions. This approach enables seamless experiment analysis as LaunchDarkly syncs experiment exposure data into the warehouse, computes metrics directly against warehouse tables, and aggregates results for reporting. Recent upgrades have introduced advanced statistical capabilities such as sequential testing, multiple comparisons correction, the ability to add metrics post-experiment start, and result segmentation, all aimed at enhancing the confidence and precision of experiment results. These developments ensure that product and data teams can conduct faster and more accurate experiments while maintaining data governance.

Code correctness in isolation doesn't guarantee production-readiness, which requires considering real-world conditions and interactions. While AI-powered Site Reliability Engineering (SRE) agents excel at detecting production issues, they often lack the capability to act on these detections, leading to a gap in incident response. The AWS DevOps Agent and LaunchDarkly MCP Server integration addresses this by using feature flags as precise levers for operational control. Feature flags allow teams to contain issues quickly, providing an immediate and reversible solution without redeployment, which is crucial in today's fast-paced, AI-driven development environments. This integration enables a seamless workflow where agents can query flags, recommend actions, and maintain an audit trail, all within the existing workflow, enhancing uptime and availability. By recommending flag configurations during release management and offering flag toggles as a first response during incidents, this approach helps teams maintain system stability and minimize customer impact.

LaunchDarkly has announced a strategic shift to utilizing New Relic for its primary observability and telemetry workloads to enhance the control and monitoring of its software delivery platform. This move is driven by the need for reliable, transparent, and scalable observability solutions, as existing providers increasingly adopt bundled, black-box models. New Relic's focus on intelligent observability aligns with LaunchDarkly's architectural principles, ensuring data integrity, scalability, and a closed-loop system for real-time adjustments and insights. By integrating New Relic, LaunchDarkly aims to maintain high system uptime and provide customers with improved stability and innovative capabilities, allowing for real-time monitoring of feature flag impacts on latency, error rates, and throughput. This transition supports LaunchDarkly's commitment to an open, interoperable ecosystem and sets a path for enterprises to combine top-tier feature management with superior observability.

A recent discussion with LaunchDarkly executives, including CEO Edith Harbaugh, CTO Cameron Etezadi, and Head of AI Marek Poliks, highlighted the challenges of managing AI agents in production environments, emphasizing the transition from traditional software development to AI-driven processes. The conversation explored how AI agents change the dynamics of software creation, as these agents can autonomously update behaviors and outputs without human intervention, moving beyond the conventional build-test-deploy-monitor-fix cycle. The need for runtime control in managing AI agents was underscored, with a demonstration of AgentControl, a tool designed to automatically correct agent behavior in production without redeploying. This approach allows teams to optimize agent performance across various metrics while maintaining customer experience integrity. LaunchDarkly’s internal project, Fairytale, revealed that integrating structure, checkpoints, and human oversight enhances the performance and speed of AI agents, challenging the notion that AI development necessitates a trade-off between velocity and safety. The company is offering an interactive workshop, Runtime Labs, for participants to explore and experiment with runtime control concepts in AI systems.

The text discusses the challenges and solutions in managing the lifecycle of machine learning (ML) models, emphasizing the importance of MLOps to ensure consistent performance, traceability, and reliability. It highlights how model performance can degrade over time due to data drift, user behavior changes, and system updates, stressing the need for a coordinated, end-to-end approach that integrates data ingestion, feature engineering, training, deployment, monitoring, and governance. Data ingestion should be treated as a primary control layer, ensuring data quality and compliance, while feature engineering must maintain consistency between training and serving environments to avoid issues like training-serving skew. The text underlines the significance of automating model training, experiment tracking, and deployment processes using CI/CD systems, with a focus on packaging models as versioned artifacts for reliable deployment. Monitoring and observability are crucial for detecting issues post-deployment, with real-time dashboards and alerts aiding in quick problem detection. A feedback loop for retraining models based on performance insights ensures continuous improvement, supported by robust governance frameworks that automate policy enforcement and maintain compliance. Tools like LaunchDarkly are presented as effective for managing runtime controls, enabling safe model rollouts and facilitating A/B testing to optimize model performance.

Retrieval-augmented generation (RAG) systems can experience performance degradation due to uncoordinated changes and a lack of explicit versioning and ownership, which complicates tracing regressions. To mitigate this, production AI pipelines should control changes explicitly, treating retrieval, prompting, evaluation, and model selection as configurable elements. This approach enables engineers, product teams, and operations to iterate safely by using versioned configurations, automated rollback, and monitoring signals. AgentControl configs support this by externalizing parameters like retrieval depth and model selection, allowing adjustments without redeployment. This configuration-driven method ensures stability and adaptability by using metrics to guide experimentation and feature rollouts, thereby minimizing risk and enabling continuous evolution. Through dynamic optimization, cost and latency can be effectively managed by routing queries based on complexity and other attributes. Continuous monitoring of key metrics allows for proactive adjustments, maintaining production reliability. Feedback loops using structured user input facilitate ongoing improvements, ensuring that changes are data-driven and low-risk. By externalizing and controlling change, RAG systems can evolve safely and intentionally in production environments.

Feature flags and feature branching are distinct but complementary strategies in software development, each serving different stages of the delivery process. Feature branching is used during development to isolate changes and ensure code quality before integration, enabling parallel development and safe testing. It involves creating separate branches for new features, which are integrated into the main codebase only when ready. In contrast, feature flags manage production-time behavior, allowing teams to control which features are visible to users after deployment. They offer flexibility in feature rollout and enable real-time modifications without redeployment, thus providing runtime control and mitigating risk by allowing features to be toggled on or off in response to user feedback or production issues. Modern software teams often utilize both approaches to enhance release velocity, manage risk, and maintain high deployment frequency without compromising on quality or user experience, effectively decoupling deployment from release.

Release management tools are essential for modern software deployment, providing a crucial control layer between continuous integration and continuous delivery (CI/CD) pipelines and end users. These tools facilitate the coordination of deployments, manage rollout scopes, and ensure quick recovery from issues, acting as a feature control plane that determines when, how, and to whom code is released. Unlike traditional methods that pushed code directly to production, modern release management tools offer features such as deployment orchestration, progressive rollouts using feature flags, environment management, rollback capabilities, and visibility through audit trails. They complement CI/CD pipelines by offering additional safety mechanisms and control, handling the mechanics of deploying code safely to production. Different tools cater to varying needs, with some focusing on feature management, others on deployment automation, and yet others on experimentation and analytics, reflecting the diverse requirements of teams working with distributed systems that demand fast rollbacks and progressive delivery.

Artificial intelligence tools, particularly large language models (LLMs), are fundamentally different from traditional software because they are probabilistic, meaning the same inputs can yield different outputs depending on various factors like temperature settings and context. This unpredictability introduces risks such as inventing facts or generating unsafe content, necessitating rigorous experimentation and evaluation processes to optimize performance, ensure safety, and manage costs. The guide emphasizes the importance of structured experimentation, which includes A/B testing, evaluating system changes with real users, and using metrics that inform actual product impact. It outlines best practices for managing AI systems, including optimizing system messages, choosing appropriate model parameters, and ensuring responsible governance and safety measures. LaunchDarkly is highlighted as a tool that facilitates AI experimentation by enabling safe, controlled rollouts and version control, allowing for continuous, data-driven improvements without the need for extensive redeployment.