July 2026 Summaries

A small team at GitLab conducted an experiment to determine if AI agents could assist in migrating part of their legacy rate-limiting system without compromising safety. The experiment was successful, demonstrating that AI agents can be effective but also highlighting areas for improvement in existing workflows. The team used GitLab, the GitLab Duo Agent Platform, and other tools to unify two rate-limiting paths into a single implementation, focusing on observability, testability, and operational consistency. The process involved a structured loop with AI agents drafting specs, implementing changes, and reviewing merge requests, while humans retained control over scope, architecture, and final reviews. The project faced challenges such as a shadow-mode miss and infrastructure constraints, which underscored the importance of human oversight and judgment. By mid-June, the migration was successfully completed for all cohorts, with plans to address the higher-volume RackAttack layer next. The experiment illustrated the value of both AI agents and the human element in achieving a successful migration.

GitLab's restricted access feature is designed to help organizations manage seat costs predictably by blocking the addition of new billable users once all licensed seats are occupied. This feature, available on both GitLab.com and Self-Managed, ensures that seat usage aligns with purchased seats without reversing existing overages, allowing organizations to avoid unexpected seat growth before renewal. Recent improvements have enhanced the feature's integration with identity providers, allowing users provisioned through systems like SAML, SCIM, or LDAP to be assigned non-billable Minimal Access roles, thereby preventing immediate overages while maintaining centralized identity management. Dormant users who reactivate are now placed in a pending approval state rather than automatically consuming a seat, and clearer warnings and notifications have been implemented to assist administrators in managing seat limits effectively. Restricted access differs from the user cap feature, as it directly ties to the number of available seats, automatically disabling user cap when activated. It provides a more controlled and transparent operational model for managing user provisioning and reactivation while helping reduce billing surprises.