TLDR DevOps 2026-09-25

Buildkite, the pro-tool that hands you building blocks for CI orchestration and then gets out of the way.

Enough with constant downtime. No more band-aids or burden.

We provide programmable primitives you co-ordinate, running only what changes, across any cloud, any chip, your secrets, your infra.

Control without compromise. Build faster, more securely, more reliably, with the CI engineered for frontier scale, used by the teams setting the pace.

You deserve better CI. Start with our 30-day trial. Kick the tires →

Google Kubernetes Engine adds Pod snapshots, capturing a workload's running CPU and GPU memory state to restore it on demand, cutting AI inference startup by up to 89 percent and loading a 70B parameter model in 37 seconds. Codeway's Retake cut startup to eight seconds using the feature.

OpenAI launched an improved prompt caching system for the GPT-6 family that delivers higher cache hit rates by default. Developers get discounts of up to 90% on cached input tokens when the system reuses shared prefixes within a 30-minute window. New tools include a Prompt Caching Dashboard to track hit rates and a diagnostics tool to investigate unexpected misses. Developers can also prewarm the cache to reduce latency or adjust reasoning effort between responses without breaking cache.

Cloudflare introduced an Agent Development Lifecycle to replace traditional software pipelines with autonomous, event-driven systems for AI coding agents, built on its Workflows product and a new @cloudflare/ci tool. It adds OpenTelemetry-based observability and an Agent Access Model that issues short-lived, capability-limited credentials to prevent lateral movement.

Datadog fine-tuned Qwen3.5-9B on traces from a larger teacher model to automate change attribution during incident investigations. The specialized student model achieved 87% of the teacher's recall while reducing inference costs by roughly 20 times to $0.003 per investigation. A single 40 GB A100 GPU can support approximately 100,000 of these investigations per week.

AI systems increasingly depend on hidden layers of human labor, including data labeling, content moderation, model evaluation, and other forms of contract work that shape how models behave. Making those labor relationships visible helps explain where AI systems get their capabilities, how failures are caught, and which costs are shifted onto workers outside the companies building the models.

Traffic skyrocketing? If you're on Microsoft Azure, there's no reason to panic. When user demand spikes, Azure adjusts capacity in the background. Performance stays steady with no manual intervention, no late-night firefighting, and no scrambling to explain why the site went down during the big launch. Build without limits with Azure

NVIDIA has open sourced its Model Optimizer library, which offers techniques like quantization, pruning, and distillation to accelerate deep learning models. The tool supports inputs from Hugging Face, PyTorch, or ONNX and exports optimized checkpoints for deployment frameworks like TensorRT-LLM and vLLM. Users can install the stable release via PyPI or use NVIDIA container images with the software pre-installed.

Strands Agents is a new open-source SDK for building AI agents in Python and TypeScript. It offers a "harness" with benchmarked defaults for models and memory, or developers can use the underlying SDK to write their own agent loops. The toolkit includes support for tools, structured output, and model providers such as Amazon Bedrock, Anthropic, and OpenAI.

The Docker Sandbox Kit Specification v3 defines how to package an agent's permissions and dependencies as an ordinary OCI image. The format specifies network rules, credentials, and volumes, allowing developer tooling to scan, sign, and diff agent authority changes. Docker Sandboxes is the first conforming runtime for the open source specification.

Google's multi-cluster GKE Inference Gateway pools globally distributed accelerator capacity behind a single endpoint, routing LLM traffic based on live KV-cache utilization rather than round-robin. A 17,000-node benchmark across three regions showed near-linear throughput scaling with less than 1 percent routing overhead and a 99.9 percent success rate.

When an alert fires on Amazon ECS, a tutorial in AWS Marketplace walks you through the trace to root cause, one path instead of five dashboards. Read the tutorial.

Amazon CloudWatch Omni is a unified observability experience for application and AI workloads.

Google is adding a persistent, encrypted memory layer to its Private AI Compute platform, letting AI assistants retain context across devices while keeping data unlockable only with device-held keys.

SemiAnalysis's ClusterMAX 3.0 report rates 77 GPU cloud providers out of 323 tracked, drawing on over 200 customer interviews and testing across audit, performance, and reliability phases.