Mastra shipped a steady stream of capabilities across agent orchestration, observability, and storage over the last ninety days. The platform consolidated observational memory (compressing conversations 5–40x with reflection-based condensing), expanded workspace tooling with AST-aware edits and background process management, and introduced structured evaluation primitives—versioned datasets with SCD-2 item history and experiments that run agents against them with configurable scorers. Authentication and RBAC landed end-to-end via pluggable auth and Okta integration, while observability storage got first-class schemas and in-memory implementations across all signals (scores, logs, feedback, metrics, discovery). On the infrastructure side, supervisor patterns enabled multi-agent coordination, dynamic model routing with fallback arrays added runtime flexibility, and workspace filesystem mounts unified access across S3, GCS, and local storage through a single path tree.