Experimentation Tools News: GrowthBook 5.0 Ships, Aug 2026

The fortnight’s theme is automation moving into the experiment itself. GrowthBook shipped a major version built around AI agents, and PostHog spent it making feature flags harder to break and its experiment workflow drivable from an assistant. Here is what changed and whether it earns a place in a small programme.

GrowthBook 5.0 ships with AI agents and a rebuilt visual editor

On 20 July GrowthBook released version 5.0, its first major release in about a year, and the headline is that AI now runs through the whole product. It ships official agent skills, an in-app assistant and a CLI that covers the entire API, alongside a rebuilt AI Visual Editor that lets you build browser-based tests with image generation and Figma support. The less flashy changes matter more day to day: Product Analytics is out of beta and generally available, feature flag governance adds guardrails that catch bad flags before they ship, and experiment setup is streamlined with scheduled starts and cheaper warehouse queries. For a small team the practical wins are the scheduled starts and the flag guardrails, because most botched tests come from a flag going live wrong or a test starting before the tracking is ready, not from a lack of clever analysis. If you are weighing GrowthBook against the field, our best A/B testing tools guide and the side-by-side comparison cover where it fits. The full write-up is on GrowthBook’s blog.

PostHog lets feature flags fall back to a safe value while loading

On 24 July PostHog shipped a batch of feature-flag reliability changes, and the most useful is a fallback value. You can now pass a defaultValue to isFeatureEnabled in the JavaScript and React Native SDKs, so a flag returns a sensible answer while the definitions are still loading rather than flickering or defaulting to off. The same release made flag dependencies more resilient, so a payload keeps evaluating when an older dependency uses a numeric key instead of blocking every other flag in the project, and cohort-based flags now treat unsupported filter conditions as a non-match rather than failing outright. This is unglamorous plumbing, but a flag that briefly reads wrong on page load is exactly the kind of thing that quietly pollutes an experiment’s early data and sends you chasing a result that was never real. Guarding against that is the same discipline our CRO best practices guide argues for before you trust any readout. The detail is in PostHog’s changelog.

PostHog adds 32 agent tools, including experiment recalculation

On 3 August PostHog added 32 new tools to its MCP server, the layer that lets an AI assistant such as Claude act on your project directly, and several are aimed at experiments. Agents can now recalculate experiment metrics, fetch experiment results and check the status of a recalculation run without anyone opening the dashboard. The appeal is turning the mechanical parts of a test, forcing a fresh metric calculation or pulling the current numbers, into something you can ask for in plain language. The risk is the same one every shortcut carries: an assistant will happily report a result before it is mature or before a sample ratio mismatch check has passed, so the judgement about whether a number is trustworthy still sits with you, not the tool. Our CRO research methods guide and the sample ratio mismatch calculator cover the checks that no amount of automation replaces. The change is listed in PostHog’s changelog.