Google just shipped two changes that should change how platform teams run agentic workloads at scale: console-driven reinforcement-learning fine-tuning jobs for Gemini models, and a deferred execution tier that queues non-latency background agent tasks to run in off-peak windows (both announced mid‑September 2026).
The reinforcement-learning fine-tuning (RLFT) preview is more than a checkbox. You can run fine-tuning jobs from the console with either Python reward functions or model-based reward models, see real-time training and evaluation metrics, and validate behavior inside Agent Studio. Those are the primitives teams need to move from ad-hoc prompt tweaks to measurable policy tuning: programmatic rewards, live metrics, and an integrated test loop.
At the same time Google adjusted lifecycle dates for older model weights, which underscores a second point: model lifecycle management is now an operational problem, not a research footnote. If your deployment and CI/CD pipelines don't account for retirement dates, you will be running unsupported models before you notice.
The deferred execution tier is the other structural change. Instead of forcing all agents into the same low-latency, on-demand path, Google now offers a preview tier that queues background, non-urgent workloads and schedules them for off-peak execution. In practice that means:
- capacity planning that separates 99th‑percentile interactive concurrency from batch agent throughput,
- an easier route to leverage cheaper, lower-priority GPUs or preemptible capacity for background scoring,
- clearer cost attribution and queueing policies for agentic workloads that are tolerant to latency.
This is the right call. Teams have been shoehorning everything into low-latency online paths, which inflates CPU/GPU baselines and messes with autoscaler behavior. Splitting interactive from deferred is what platform engineering should have done two years ago.
But it also changes your platform checklist. Treat agent workloads like stateful ML infra: provision separate node pools (GPU/non‑GPU), enforce QoS and pod priority classes, and bake queuing and retry semantics into your orchestration layer. Expect to add model-version-aware admission controls and billing tags to avoid accidental execution of retired models during deferred runs.
Other mid‑September updates back this operational shift. Artifact Registry added support for Conda packages (preview), which helps you produce reproducible Python environments for agent runtimes without exploding container image permutations. Google also expanded confidential computing options to include GPU configurations (including H100 in some offerings), giving teams a way to protect model weights and inference while staying on Google Cloud. The Architecture Center published a prescriptive SQL Server Always On availability group reference with regional failover for transactional backends agents may depend on.
Operationally, Google’s release notes point one direction: GKE plus unified ops tooling as the platform for mixed agentic and traditional services. Expect more features to smooth mixed workloads — from node autoscaling tuned for bursty agent traffic to observability that joins model telemetry with typical service metrics. If you followed the earlier Vertex AI and Gemini previews, this continues the same trajectory; changes to GKE defaults and model-routing gateways make the technical path explicit.
There’s one real risk: agents increase your attack surface. RLFT runs that accept Python reward functions and deferred tiers that run queued jobs on cheaper hardware both widen trust boundaries. Lock down who can submit fine-tuning jobs, enforce least privilege on model artifacts and Conda packages, and audit any execution that touches confidential models.
Final point: this is where platform engineering pays off. If you treat agents as first-class—separate scheduling tiers, enforceable model lifecycles, and reproducible runtime environments—you'll cut costs and avoid surprises. If not, next month’s bill and a pair of retired-model incidents will teach you the hard way.
Sources
- Gemini Enterprise Agent Platform release notes – September 15, 2026
- Google Cloud Release Notes – September 16, 2026
- What's new in the Architecture Center
- Forrester Wave Public Cloud Platforms Q3 2026 report (Google Cloud blog, GKE and agent workloads)
- Google Cloud latest news and announcements (Sept 10, 2026 rollup)