BRIEF #13
July 13, 2026

Platform Pulse: Open Knowledge, Agent Sandboxes, and Elastic Training

This week's engineering brief highlights Google Cloud's continued acceleration into the agentic era. We review the new Open Knowledge Format, examine the public preview of secure Cloud Run sandboxes for untrusted agent code, and explore robust elastic training architectures for TPUs.

🤖 AI Agents & Language Models

The ecosystem is actively standardising agentic workflows, moving from static prompts to code-first frameworks equipped with graph-based orchestrations and robust testing flywheels.

  1. AlphaEvolve GA: AlphaEvolve, a code optimisation and discovery agent built on Gemini, is now generally available to help solve complex business and research problems.
  2. Gemini Omni Flash and Nano Banana 2 Lite: Nano Banana 2 Lite is generally available for everyone, and Gemini Omni Flash is available in public preview within the Gemini Enterprise Agent Platform.
  3. Claude Apps Gateway: This self-hosted service sits directly between local Claude Code clients and the Gemini Enterprise Agent Platform to ensure secure deployments on Google Cloud.
  4. Google ADK 2.0: Google has detailed the rationale, new features, and migration paths for the newly stable Agent Development Kit 2.0.
  5. ADK Go 2.0: The latest release introduces a graph-based workflow engine, built-in human-in-the-loop orchestration, and automated resilience mechanisms like exponential backoff retries.
  6. Polymorphic Multi-Agent Systems: Developers can build scale-proof multi-agent architectures that solve context bloat and attention diffusion through dynamic, metadata-driven validation.
  7. Genkit Agents API: Available in preview for TypeScript and Go, this API packages message history, tool loops, and streaming into a single interface for conversational AI.
  8. Agent Quality Flywheel: A new developer skill automates a five-stage evaluation flywheel to safely validate prompt tweaks and prevent widespread production regressions.
  9. Web Manual-like QA with Gemini: Gemini's "Computer Use" capability visually perceives and interacts with web pages, automating QA testing directly from screen pixels to withstand UI changes.
  10. DiffusionGemma: An experimental text-generation model using diffusion-based parallel generation instead of autoregression to process 256-token blocks simultaneously.

📊 Data Analytics, BigQuery & Storage

Data engineering practices are pivoting toward semantic intelligence, releasing Open Knowledge specifications and pushing new per-second FinOps billing paradigms across BigQuery.

  1. Open Knowledge Format (OKF): The proposed OKF standard represents data as code, creating a portable digital map to standardise documentation and solve scattered knowledge challenges for AI agents.
  2. Conversational Analytics in BigQuery: Now generally available, this feature allows users to query data and generate reports using natural language across BigQuery and Lakehouse.
  3. AlloyDB AI Functions: New performance boosts and cost savings have been introduced for AlloyDB AI functions, bringing Gemini's intelligence closer to the data.
  4. BigQuery Fluid Scaling: BigQuery has eliminated the previous 60-second minimum for reservation billing, shifting to per-second consumption to reduce cooldown waste on bursty workloads.
  5. BigQuery AI.AGG Function: This new function allows data teams to synthesise the big picture and analyse trends across unstructured data logs and documents at scale.
  6. E-Commerce Churn Prediction with TabFM: Google's TabFM foundation model enables zero-shot tabular predictions in BigQuery, leveraging in-context learning via SQL without prior training.
  7. Dataform Incremental Load Optimisations: Engineers have identified four practical techniques to drastically reduce execution times and BigQuery costs for production Dataform pipelines.
  8. BigQuery Information Schema Queries: Five essential INFORMATION_SCHEMA queries that data engineers must run to correctly target BigQuery performance optimisation efforts.

🔐 Zero-Trust Security & Threat Intelligence

Organisations are hardening architectures to fight sophisticated extraction tactics and proxy networking abuse, driving rapid adoption of internal AI threat defences.

  1. Confidential AI with Apple: Google Cloud is expanding its Confidential Computing capabilities by collaborating with Apple on its Private Cloud Compute (PCC) systems.
  2. Dutch DPIA Approval for EU Public Sector: Google Cloud has confirmed its commitment to data protection with a Dutch DPIA approval, offering a safer choice for European public sector organisations.
  3. Google Threat Intelligence and Wiz ASM: Google Threat Intelligence has integrated with Wiz Attack Surface Management to help match real-world exposures against real-time adversary activity.
  4. Detecting AI-Powered Threats: Google Security Operations now works alongside AI Threat Defense to monitor and contain autonomous threats, particularly from unpatched code.
  5. Active ADFS Signing Keys Recovery: Threat intelligence reports detail how adversaries are exploiting Machine DPAPI to forge high-privilege SAML tokens.
  6. Cisco Switch Syslog Ingestion: A lab guide outlining how to use a headless Raspberry Pi 5 to sanitise Cisco IOS headers and bridge legacy network hardware with Google SecOps.
  7. Disruption of Malicious Proxies: Google successfully disrupted the NetNut malicious residential proxy network, safeguarding over 2 million hijacked consumer devices.
  8. STOCKSTAY Intelligence Gathering: Analysts have uncovered a new backdoor named STOCKSTAY within the pro-Russia influence ecosystem, continually developed by the threat actor Turla.

⚡ Cloud-Native Infrastructure & Serverless

To run high-compute AI boundaries securely, infrastructure teams are rapidly provisioning hardware queues, sandboxing runtime logic, and migrating towards elastic TPU clusters.

  1. C4N Machine Series GA: The C4N family is now generally available, delivering the highest per-vCPU network and block storage I/O performance for x86 enterprise applications.
  2. Cloud Run Sandboxes in Preview: Announced at WeAreDevelopers World Congress, Cloud Run sandboxes offer a native, ultra-fast runtime built specifically for executing untrusted AI agent code.
  3. Kueue in Google Cloud Console: Google has integrated the Kueue standard directly into the Cloud Console, providing a visual interface for managing high-performance GPUs and TPUs.
  4. Cloud CDN Private Bucket Access: Cloud CDN and external ALBs now support private bucket access via IAM-only permissions, eliminating the need to expose Cloud Storage buckets or generate signed URLs.
  5. 100x SRE: Agentic GKE Capacity Optimiser: Engineers can build an autonomous agent using Google Antigravity 2.0 to dynamically generate optimal GKE manifests and resolve "Pending Pods".
  6. 100x Platform Engineering with Fabric FAST: Antigravity 2.0 can now autonomously architect, validate, and self-heal enterprise-grade GCP landing zones using data-driven "Recipe" Skills.
  7. Cloud Run CPU-Throttling Mystery: A deep dive explaining how request-based allocation mode on Cloud Run freezes background tasks, requiring developers to switch to instance-based CPU allocation.
  8. Scaling LLM Inference with Managed Lustre: Enterprise LLM inference on GKE can now offload KV caches to Google Cloud Managed Lustre for better long-context support and reduced GPU costs.
  9. Elastic Training with MaxText: Google's JAX ecosystem uses Pathways to convert TPU hardware failures into Python exceptions, allowing distributed AI training jobs to automatically replace broken workers and resume in seconds.
0

From the Community

No community links this week.

Enjoyed this brief?

Don't miss the next drop.