January 2026 Summaries

Firebase Crashlytics can now export fatal crashes, non-fatal errors, and ANRs to Google Cloud Logging, connecting mobile stability data with Google Cloud’s observability and monitoring tools. The integration enables teams to create custom log-based metrics and alerts with tailored thresholds, queries, notification channels, and regional or error-specific conditions, helping reduce unnecessary alerts. Developers can investigate crash data through Logging Query Language or SQL, filter events by device details and app-defined keys, and use Gemini in the Cloud console for query assistance. Cloud Monitoring dashboards can combine crash-free rates, session volumes, error counts, and critical user journey metrics, while optional Firebase Sessions exports provide additional context. By placing client crash reports alongside backend logs, the feature also supports full-stack debugging and correlation of device failures with server-side activity. Setup is available through Firebase Project Settings under Integrations, with Cloud Logging offering a 50 GB monthly free tier per project before standard pricing applies.

As of August 2026, Vertex AI has been renamed Agent Platform, and its Prompt Optimizer is presented as a data-driven tool for refining repeatable AI tasks by using examples of known inputs and ground-truth outputs to iteratively improve prompt instructions and evaluate results. The Firebase team tested it for generating YouTube video descriptions from scripts, assembling historical scripts and human-written descriptions in a CSV file stored in Google Cloud Storage, then using a Colab Enterprise notebook to optimize an initial system prompt. With Gemini 2.5 Flash, instruction-and-demo optimization, and evaluation metrics, the service ran a custom training job that compared prompt variations against an original baseline; processing 45 examples took about four hours and consumed millions of tokens. The optimized prompt added detailed constraints for summaries, timestamps, resources, speakers, featured products, and hashtags, producing descriptions that were more consistent with human-authored examples and less variable than the original outputs. Although the process cost several hundred dollars, the example concludes that its improved consistency and reduced need for human review justified the expense for this workflow, while noting that users should weigh optimization costs against expected benefits.

Google Cloud and Firestore have introduced a redesigned query engine for Firestore Enterprise edition, adding Pipeline operations and more than 100 query capabilities that enable multi-stage transformations, arbitrary aggregations, array unnesting, map operations, and regular-expression string matching. Pipeline operations allow developers to compose queries from stages, expressions, and functions, illustrated by a recipe application that extracts and ranks popular tags directly from document arrays without maintaining separate metadata. Enterprise edition makes indexes optional rather than automatically indexing collections, potentially improving write performance and reducing storage costs, while tools such as Query Explain and Query Insights help developers identify when indexes are needed; it also supports sparse, non-sparse, and unique indexes. Available now for Android, iOS, web, and Admin SDKs, with Flutter, Unity, and C++ support planned, the feature requires a new Enterprise database and updated SDKs. Existing Standard edition users may migrate data through Firestore import and export tools but must recreate indexes and security rules, while Standard edition and its existing query methods will remain supported indefinitely.

Continuous improvement is presented as an AI-agent strategy in which an initial model response is repeatedly reviewed and refined by another model, or informed by a human-in-the-loop intake process, to produce more reliable and tailored results than a single prompt. Using Google’s ADK as an example, the text describes a workout-planning system where a reception agent first gathers details about a user’s goals, equipment, experience, and physical limitations, then passes the information to a LoopAgent containing a workout-design agent and a critical-review agent. In the example, this process replaces an unsuitable generic arm-focused workout with a plan that accounts for lower-back pain, limited equipment, recovery needs, and balanced shoulder development. The approach can improve response quality without requiring an excessively detailed initial prompt, but it increases latency and model costs because multiple iterations are needed, making it most appropriate for high-value, infrequent tasks such as personalized plans rather than routine requests.

Firebase App Hosting now supports Turborepo in addition to Nx, expanding its ability to deploy JavaScript and TypeScript monorepos that contain multiple applications and shared libraries. Turborepo uses a root-level turbo.json configuration to define tasks and their dependencies, creating a dependency graph that ensures shared packages are built before dependent applications while running unrelated tasks in parallel to improve build speed. When deploying a Turborepo, App Hosting automatically analyzes the repository’s turbo.json, builds required dependencies, and deploys only the selected target application. Users can specify the target app directory during GitHub repository import or Firebase CLI initialization and later modify it in firebase.json for local-source deployments. Remote caching is not yet supported, but the integration is intended to make App Hosting’s continuous deployment pipeline more efficient for large, complex monorepos.