June 2026 Summaries

Squarespace's launch of a blank template for their website-building platform, aimed at providing users with complete creative freedom, initially drew more users into their CMS but failed to convert them, revealing significant insights about their customer base. Lina Blackman, Director of Product Analytics at Squarespace, discussed how the experiment highlighted the divide between technical users seeking customization and small business owners needing guided design assistance. This misstep led to the development of Blueprint, an AI-guided builder that better serves their varied user demographics. Blackman emphasizes the importance of treating failed experiments as valuable data rather than losses, advocating for a disciplined approach that identifies granular user segments and reevaluates the viability of ideas. With Squarespace running numerous experiments annually, the focus is on understanding and learning from failures, which informs future product decisions and helps prevent potential business detriments. As AI becomes more integrated into the testing process, Blackman stresses the need for a balance between automation and human reasoning to enhance product development and user experience.

GrowthBook has introduced an open-source agent skills library designed to enhance the functionality of AI coding agents like Claude and Cursor by providing them with structured playbooks for executing tasks such as creating feature flags and analyzing experiments. Instead of relying on the agent's guesswork, these skills, written in Markdown, guide the agent through best practices for specific GrowthBook tasks, ensuring consistency and accuracy. The skills utilize the GrowthBook REST API and follow the Agent Skills standard, making them compatible with various AI agents. By encoding procedures into steps that the agent consistently follows, the skills prevent errors such as incorrect flag rollouts or incomplete experiment analyses, thus adding a layer of reliability to AI-driven processes. Users can access, adapt, and contribute to the skills library on GitHub, fostering a community-driven approach to refining and expanding this toolset.

Flagsmith is an open-source feature-flagging and remote-configuration platform favored by many engineering teams due to its bootstrapped and self-hostable nature. However, as teams mature and require more advanced capabilities like automated rollbacks and comprehensive diagnostic workflows, Flagsmith's limited resources and features compared to larger competitors can become a limitation. The platform lacks AI-native flag workflows, comprehensive integrations with data warehouses for analytics, advanced statistical engines for experimentation, and robust pricing models, which can lead teams to seek alternatives. Flagsmith offers flexibility through its cloud-based and self-hosted deployment options, making it suitable for solo developers and mid-market teams in regulated industries. But for teams needing more sophisticated feature flagging, experimentation, and integration capabilities, alternatives like GrowthBook, LaunchDarkly, Statsig, PostHog, Unleash, Split by Harness, and Eppo offer varied strengths catering to different needs, such as strong AI-native features, enterprise governance, and tight integration with other platforms.

Unleash is a well-established open-source feature flag platform launched in 2015, primarily catering to enterprise engineering teams by providing a self-hosted, API-first method for toggling features in production. It is trusted by companies like Visa and Deutsche Telekom but faces limitations as feature flagging evolves, such as the lack of real-time flag propagation, server-side identity storage, and AI-native workflows. The open-source version will be deprecated by December 2026, pushing users towards the enterprise version or alternatives. Unleash excels in governance features like change request approvals and granular RBAC, making it suitable for regulated industries. However, its polling-based architecture, lack of experimentation tools, and high costs for non-engineering access lead some teams to explore other feature flagging platforms. Alternative platforms like GrowthBook, LaunchDarkly, and Flagsmith offer different strengths, such as warehouse-native experimentation, real-time flag updates, or bundled analytics, catering to teams seeking advanced capabilities or integrated solutions.

In a rapidly evolving landscape of AI-driven software development, Gene Kim highlights the challenges of maintaining system stability by referencing control theory's Nyquist stability criterion, which suggests that the control system must be faster than the governed system. The 2025 DORA report reveals that while AI adoption in software engineering is widespread, the accompanying control mechanisms are lagging, leading to increased incidents and complexities in managing AI-generated code. Traditional development models, such as CI/CD and QA, struggle with the probabilistic nature of AI, necessitating new approaches like feature flags to enable controlled rollouts, experimentation, and observability. Feature flags allow developers to manage AI-induced risks by gradually introducing changes, enabling instant rollbacks, and controlling exposure to minimize potential disruptions. The integration of feature flags into AI workflows not only mitigates risks but also facilitates regular testing, model comparisons, and the seamless transition between AI and non-AI paths. As AI continues to transform software development, platforms like GrowthBook offer solutions for implementing effective control layers, ensuring reliable and scalable AI deployment by leveraging warehouse-native architectures and experimentation engines.

In 2026, A/B testing platforms have evolved to become essential tools for organizations seeking to innovate through experimentation-driven development, with applications now extending beyond marketing teams to entire product teams. The text discusses several key platforms and their unique offerings: GrowthBook stands out for its warehouse-native integration, open-source options, and developer-first approach, making it ideal for organizations focused on scalability and compliance; Statsig offers a straightforward, all-in-one solution but may not meet the needs of sophisticated data teams due to its limited flexibility and rising costs; Eppo provides statistical precision and is well-suited for data-driven teams, although its feature flagging capabilities may not be comprehensive; and legacy platforms such as LaunchDarkly, Optimizely, and Adobe Target face challenges in cost, integration, and flexibility, making them less appealing for modern workflows. The guide underscores the importance of selecting the right platform by aligning organizational priorities with platform capabilities to enhance the ability to iterate, scale, and innovate effectively.

Feature flags, while useful for separating deployment from release, can inadvertently lead to technical debt when not managed properly. Over time, unused or stale flags accumulate, creating risks such as increased cognitive load, security vulnerabilities, and operational hazards. This debt often goes unaddressed due to a lack of ownership, fear of removal, inadequate lifecycle policies, and insufficient flag management practices. Despite 77% of developers intending to remove unnecessary flags, audits reveal that 75% of toggles remain long after their purpose is fulfilled. Automated tools and practices, such as stale detection and code references, are essential for identifying and managing these flags, yet organizational pressure and a lack of clear governance often impede effective cleanup efforts. GrowthBook offers solutions by providing automated stale detection, code mapping, and approval workflows to help manage and reduce flag debt, emphasizing the importance of integrating flag removal into the software development lifecycle to prevent unnecessary technical debt.

Feature flags are a powerful tool that enhance the CI/CD process by separating deployment from release, allowing for more controlled and measured rollouts of new features. They address the limitations of traditional CI/CD approaches, which often treat deployments as all-or-nothing events, potentially leading to costly downtime and reactive fixes when issues arise. By using feature flags, teams can deploy code continuously while controlling its visibility to users, enabling gradual rollouts, targeted testing, and immediate rollbacks without redeploying code. This approach not only increases the frequency and confidence of deployments but also reduces incident costs and improves overall business outcomes. Feature flags facilitate continuous experimentation by embedding measurement into every release, allowing teams to make data-driven decisions rather than relying on intuition. The integration of feature flags with CI/CD pipelines, as demonstrated by platforms like GrowthBook, provides automated rollout schedules, guardrail monitoring, and rollback capabilities, ensuring that deployments are safe and effective.

Geoffrey Bell, an experimentation product specialist at Ford Credit, emphasizes the importance of rigorous testing before implementing new features, as illustrated by an experiment involving a vehicle selector on Ford Credit's online prequalification form. Initially, the hypothesis was that allowing prospective buyers to choose their desired vehicle early in the process would increase completion rates, but the test revealed the opposite, with completion rates declining. This outcome highlighted the value of experimentation in preventing costly mistakes and underscored the need for strategic placement of features. By repositioning the vehicle selector to appear after prequalification, Ford Credit achieved better results, demonstrating that timing and context are crucial in the customer journey. Bell advocates for a balanced approach to experimentation, where both the gains from successful tests and the savings from avoided poor implementations are considered, reinforcing the notion that a prevented error is as valuable as a successful outcome. This approach not only builds credibility but also aligns customer behavior insights with business metrics, ultimately enhancing decision-making and driving meaningful revenue impact.

In A/B testing, metrics or Key Performance Indicators (KPIs) are crucial for evaluating experiments' success and guiding business decisions. These metrics, which include binary, continuous, and ratio types, capture various aspects of user behavior and help in assessing the impact of interventions. Binary metrics simplify behavior into yes-or-no outcomes, while continuous metrics provide detailed insights into user engagement and intensity. Ratio metrics, such as average revenue per paying user, focus on the relationship between two variables. Each metric type requires specific statistical methods for analysis, and understanding these can significantly affect the interpretation of results. KPIs are categorized into primary, secondary, and guardrail metrics to ensure focused and transparent experimentation. Primary metrics define success, secondary metrics provide context, and guardrail metrics monitor for unintended harm. The choice of a primary KPI should balance business relevance and statistical efficiency, often requiring careful consideration of trade-offs and the potential use of proxy or derived metrics. Understanding and strategically deploying the right KPIs transform experimentation from guesswork into a systematic decision-making process that aligns with business goals.