🔥 Top Story
AI’s best coding agent fails 60% of the time — and the data backs it up
The Real-SWE benchmark, built by Y Combinator-backed Specific Labs, tests coding agents against private production codebases from real companies instead of public-repo problems. Fable 5.1 running on Claude Code topped the leaderboard with a 38.8% success rate, followed by GPT-6 Astra on Codex CLI at 33.8% and Gemini 3.8 Flash on Gemini CLI at 31.2%. That means even the best-performing agent fails more than six times out of ten on unfamiliar private code. Failures were mostly traced to missed requirements, integration errors, and unverified assumptions rather than raw coding ability. The results suggest that scores on public coding benchmarks don't necessarily translate to navigating an unfamiliar production codebase. The piece frames this as a reason to weigh benchmarks reflecting your own codebase over headline leaderboard numbers when picking a coding agent.
💡 Why it matters: Platform teams rolling coding agents into CI/PR workflows should not trust public leaderboard scores as-is, and instead validate failure rates and failure modes (missed requirements, integration errors) against samples of their own codebase.
🔗 Read more · The New Stack
Kubernetes & Cloud Native
Kubernetes Changed Block Tracking API - Beta Differences
Kubernetes
Changed Block Tracking (CBT) support for CSI drivers first shipped as Alpha in September 2025, and graduated to Beta alongside the March 2026 v1.0.0 release of the external-snapshot-metadata project. The headline change in this Beta is that the SnapshotMetadataService CRD was promoted from v1alpha1 straight to v1beta1, with the old alpha version removed rather than served side-by-side, so operators must go through a one-time manual upgrade. CBT itself is an optional API that lets CSI drivers securely expose metadata about a volume's allocated blocks, or the changed blocks between two CSI VolumeSnapshot objects of the same PersistentVolume, enabling efficient differential backups. The current scope is limited to block volumes; changed-list tracking for file volumes or network file shares is explicitly out of scope for now. For the rest of the Beta cycle, the focus is on broader CSI driver adoption and gathering operational feedback ahead of a GA push. The post was published on the official Kubernetes blog on September 14, 2026.
💡 Clusters already using alpha CBT for differential backups need a planned migration before upgrading, since the SnapshotMetadataService CRD's alpha version is dropped rather than served alongside v1beta1, which can silently break backup pipelines.
Kubernetes v1.37: Memory QoS Graduates to Beta
Kubernetes
Kubernetes v1.37 promotes Memory QoS to Beta and turns it on by default on Linux nodes running cgroup v2, where it uses the memory controller to give the kernel better guidance on how to treat container memory. The feature was first introduced as Alpha back in v1.22 and was expanded in v1.36 with tiered memory reservation. With the Beta promotion, every v1.37 kubelet has the MemoryQoS feature gate enabled without any configuration change, but it's safe by default because the stock kubelet configuration doesn't enable memory throttling or reservation, so no memory.high, memory.min, or memory.low values get written to cgroups unless explicitly configured. The key behavioral change is that the memoryThrottlingFactor default flipped from 0.9 to null, meaning memory.high is no longer set automatically, which preserves backward compatibility across upgrades. Cluster operators can opt into memory throttling, tiered reservation via memoryReservationPolicy, both, or disable the feature entirely through explicit kubelet configuration. The change ships as part of the broader Kubernetes v1.37 release.
💡 Since MemoryQoS ships enabled by default but stays a no-op until explicitly configured, operators upgrading to v1.37 have a good window to deliberately tune memory.high and memoryReservationPolicy per workload instead of relying on blunt OOM kills.
Cilium 1.20: Gateway API ExternalAuth, TCPRoute/UDPRoute, ENI IPAM for IPv6, and more
CNCF
The CNCF blog introduces the Cilium 1.20 release. Gateway API support jumps from v1.4 to v1.6, adding an ExternalAuth filter (based on GEP-1494), CORS filters, ListenerSets, and TCPRoute/UDPRoute support for non-HTTP traffic, so the same Gateway API already used for HTTP or gRPC can now manage L4 services too. The ExternalAuth filter lets HTTPRoute requests be authenticated and authorized through an external service before they reach the application. On the IPAM side, AWS ENI IPAM mode gains beta support for IPv6, where the operator attaches an IPv6 /80 prefix to each node's ENI via Prefix Delegation and the agent assigns pod addresses out of that range. The release also introduces extensible eBPF datapath plugins, developed with Google's involvement, which the post frames as turning Cilium from a sealed networking appliance into more of a network operating system — a stable core that cloud providers can extend with their own eBPF programs independently of Cilium's release cycle.
💡 Now that Gateway API covers TCPRoute/UDPRoute as well, clusters juggling separate L4 load-balancer configs alongside Gateway API for L7 should consider consolidating onto a single set of Gateway resources when upgrading to Cilium 1.20.
AI & ML
Watch astronaut Christina Koch and Google’s James Manyika discuss space, technology, and discovery.
Google AI
In the latest installment of Google's 'Dialogues on Technology and Society' series, NASA astronaut Christina Koch sat down with James Manyika, Google's Senior Vice President of Research, Labs, Technology & Society. Koch spent 328 days aboard the International Space Station, took part in history's first all-female spacewalk, and is slated to fly around the Moon on NASA's Artemis II mission. The conversation covers viewing Earth as an electric-blue 'lifeboat' from 250,000 miles away, and how deep space exploration, AI, and societal resilience intersect. It highlights how robotics and machine learning have become foundational to long-duration missions, hazard mitigation, and extravehicular task efficiency for astronaut crews. Koch also raises the question 'are we alone,' and offers advice to future explorers to do what scares them and support the people around them. The dialogue was published on Google's blog on September 14, 2026, alongside a companion YouTube video.
💡 While not directly operational content, the human-plus-autonomous-systems model described for long-duration space missions echoes the broader shift toward AI agents and observability tooling assisting rather than replacing on-call human operators.
DevFest is back
Google AI
Google Developer Groups' community-led DevFest 2026 season runs from October 1 through December 31, 2026, spanning more than 800 localized events across 115 countries and expecting close to one million developers and builders worldwide, making it billed as the largest community-led tech conference series. The season's theme is 'Build, Secure, Scale: Developers and Builders in the Agentic Era,' with an emphasis on hands-on, practical engagement rather than passive talks. Attendees get direct exposure to Google's stack — Gemini, Google AI Studio, Google Antigravity, Google Cloud, Firebase, Android, Flutter, Angular, and Web MCP — through live codelabs, technical workshops, and agent-athons. Each GDG chapter is backed by Google but curates its own agenda, tailoring the event to its local tech ecosystem's specific needs. The season is positioned as a global on-ramp for developers to get practical experience building agentic applications on Google's tools.
💡 With hands-on codelabs on the latest agentic tooling like Google Antigravity and Web MCP across 800+ local events, platform teams should check their local GDG's schedule as a low-cost way to upskill engineers or source talent.
How Fyxer built an AI executive assistant people trust
OpenAI
OpenAI's blog profiles how email-assistant startup Fyxer built an AI executive assistant that users trust. Fyxer builds on OpenAI models with a combination of supervised fine-tuning (SFT) and Low-Rank Adaptation (LoRA) to create task-specific model variants while keeping training costs in check, and continuously improves draft quality using Direct Preference Optimization (DPO) derived from how users edit AI-generated drafts, validating each change through A/B testing. The system draws on more than 500,000 hours of annotated human executive-assistant workflow data, and splits work across 30-50 specialized models handling classification, intent prediction, memory retrieval, and draft generation rather than relying on one large generative model. The company reports that 53% of AI-drafted emails are accepted as-written, 90-day user retention sits at 90%, and revenue grew from $1M to $32M in annual recurring revenue over 2025. The system keeps improving post-deployment through ongoing real user feedback rather than a fixed, one-time training pass. OpenAI presents Fyxer as a case study in how fine-tuning and memory-based personalization layered on top of general-purpose models can be pushed toward the tone and habits of a specific individual rather than staying generic.
💡 Fyxer's architecture — dozens of small specialized models plus DPO derived from real user edits, rather than one large model — driving a 53% as-written acceptance rate is evidence that teams building internal generative AI features should prioritize task decomposition and real feedback loops over a single large model.
Cloud Updates
Agent-ready analytics: Unlocking insights with BigQuery augmented analytics
Google Cloud
BigQuery has added six new augmented analytics Table-Valued Functions (TVFs) that combine AI, ML, and statistical methods to automate insight discovery and pattern explanation directly inside BigQuery. AI.KEY_DRIVERS identifies the top drivers behind a metric's change across time periods or groups, AI.CAUSAL_EFFECT quantifies the impact of an action or event using counterfactual baselines, and ML.CORRELATION, ML.DETECT_CHANGE_POINTS, ML.TREND, and ML.SEASONALITY respectively cover relationship strength, structural shift points, underlying trend, and recurring cycles. The functions are chainable, so one function's output can feed the next analytical step, and they scale across millions of individual time series. A worked example uses Austin's bikeshare data: ML.DETECT_CHANGE_POINTS flagged a structural shift on February 11, 2018, AI.KEY_DRIVERS traced it to a 7,167.1% surge in UT Student Memberships, and AI.CAUSAL_EFFECT used ARIMA_PLUS-based counterfactual modeling to calculate a 358% volume increase, or 89,775 incremental trips. All six functions are also exposed through BigQuery's Conversational Analytics for natural-language queries, and their structured SQL output is designed to plug directly into AI agents as callable skills.
💡 Because these TVFs are chainable end-to-end from change-point detection to root-cause attribution in pure SQL, teams can embed metric-anomaly investigation directly into BigQuery scheduled queries or alerting pipelines instead of relying on a separate BI tool.
Announcing Pause/Resume and NVIDIA RTX PRO 6000 Blackwell GPU support in Dataflow
Google Cloud
Google Cloud announced general availability of Pause/Resume for Dataflow batch jobs, alongside support for the NVIDIA RTX PRO 6000 Blackwell GPU. Pause/Resume lets a failed long-running batch job resume instead of restarting from scratch, and also enables dynamically reallocating GPU or TPU capacity from lower-priority to higher-priority workloads, aimed at recovering wasted compute and boosting productivity on jobs that run for multiple days. The NVIDIA RTX PRO 6000 Blackwell GPU is available on G4 VMs with 96GB of vGPU memory and 1.6 TB/s of bandwidth, delivering a significant performance jump over the previous L4 GPU and enough headroom to run inference on 70B+ parameter models. Native Dataflow ML capabilities — RunInference, right fitting, and GPU-enabled autoscaling — run directly on this GPU, letting teams do AI inference inside Dataflow jobs without managing separate inference infrastructure. The announcement was published on September 14, 2026 by Google Cloud product manager Efesa Origbo and software engineer Danny McCormick.
💡 Combining GA'd Pause/Resume with dynamic GPU/TPU reallocation cuts both restart costs on failed multi-day batch jobs and idle-accelerator spend, so teams running large batch pipelines should revisit priority-based preemption strategies.
Google is a leader in The Forrester Wave™: Public Cloud Platforms, Q3 2026
Google Cloud
Google Cloud was named a Leader in The Forrester Wave: Public Cloud Platforms, Q3 2026, earning the highest score in the 'current offering' category and a perfect 5-out-of-5 rating in 23 of the report's 30 evaluation criteria, topping categories including vision, innovation, AI development services, database services, analytics services, and containers and Kubernetes services. Google frames its edge as an integrated AI stack built from 'silicon to systems to models,' spanning TPUs, its transformer architecture, Kubernetes, Axion processors, and Gemini models. On agent execution specifically, GKE Agent Sandbox (now GA) and Cloud Run Sandboxes (preview) run untrusted agent code under default-deny security while provisioning sandboxes in under one second at up to 300 per second, GKE Pod Snapshots serialize container memory to Cloud Storage to cut idle compute costs by up to 90% (roughly 100ms to suspend, 280ms to resume), and GKE Inference Gateway uses predictive routing to cut time-to-first-token (TTFT) by up to 70%. Forrester's assessment credits Google Cloud's vision of an 'agentic enterprise' and its AI-permeated platform as a fit for enterprises wanting rapid technology innovation. The Q3 2026 report evaluates the ten most significant public cloud providers across those 30 criteria, and Google Cloud's highest overall score put it ahead of the rest of that field.
💡 The concrete numbers behind GKE Pod Snapshots (up to 90% idle compute cost cut) and Inference Gateway (up to 70% lower TTFT) make them worth evaluating first for teams already running agent workloads on GKE.
Red Hat is named a Leader in IDC MarketScape: Worldwide Private and Hybrid Cloud Management with Automation
Red Hat
Red Hat announced it was named a Leader in the IDC MarketScape: Worldwide Private and Hybrid Cloud Management with Automation 2026 Vendor Assessment (Doc #US54644626e, June 2026). The original article could not be retrieved, so specifics such as which Red Hat product line (e.g., Ansible Automation Platform) was assessed, what particular strengths IDC cited relative to competitors, or Red Hat's exact position within the Leaders category are not confirmed here. Based on the excerpt alone, what can be confirmed is that this is a 2026 vendor assessment specifically covering the worldwide private and hybrid cloud management with automation market. Being named a Leader in this category is a signal of Red Hat's positioning in the hybrid cloud automation space, though the underlying evidence for that ranking isn't available from the excerpt. This summary is based only on the title and excerpt because the original article could not be fetched.
💡 A vendor's 'Leader' announcement from a third-party analyst report shouldn't drive procurement decisions on its own; teams should pull the actual assessment criteria and per-vendor scores from the source report before factoring it into a purchasing decision.
Modernizing Microsoft SQL Server: Choosing the right path with Red Hat
Red Hat
The Red Hat blog post, 'Modernizing Microsoft SQL Server: Choosing the right path with Red Hat,' opens with the premise that modernization is often framed as a single destination — move to containers, adopt Kubernetes, go cloud-native — when organizational reality and appetite are usually more varied than that. The original article could not be fetched, so the specific paths Red Hat actually lays out (for example, running SQL Server on RHEL versus containerizing on OpenShift, specific migration tooling, or partnership details) are not confirmed here. General Red Hat material found separately covers SQL Server being certified on RHEL and deployable on-premises, on Azure VMs, or in OpenShift containers, with Ansible Automation Platform able to automate install and management, but there's no way to confirm that content matches this specific September 14, 2026 post's actual claims. To avoid inventing unconfirmed specifics, this summary sticks to what the excerpt itself states. This summary is based only on the title and excerpt because the original article could not be fetched.
💡 Before treating containers-and-Kubernetes as the only valid SQL Server modernization path, teams should compare the actual migration risk and licensing cost differences across whatever alternative paths the vendor lays out, using the primary source rather than a secondhand summary.
From fine-tuned model to cheaper and faster inference: Speculator training on Red Hat OpenShift AI with Kubeflow
Red Hat
Red Hat's developer blog covers speculator training on Red Hat OpenShift AI with Kubeflow to make fine-tuned LLM inference cheaper and faster. Using speculative decoding, the approach can cut inference costs by up to 3x, with a workflow spanning offline data generation, draft-model training for both dense and mixture-of-experts (MoE) architectures, and serialization into a HuggingFace-compatible format that deploys directly into vLLM. Red Hat maintains a growing collection of pre-trained speculator models published on Hugging Face, covering model families commonly deployed in production today, including Llama 3.1 and 3.3, the full Qwen3 family, gpt-oss at both 20B and 120B parameter sizes, and gemma 4 at 31B and 26B. Native integration of this training and fine-tuning flow into Red Hat OpenShift AI itself is on the roadmap for later this year, which would bring the same workflow into a managed platform experience rather than requiring manual pipeline setup. The post frames this as a practical way for teams already running fine-tuned models in production to cut serving costs without sacrificing output quality.
💡 Teams already serving fine-tuned 70B-class models on vLLM should try Red Hat's pre-trained speculators for Llama, Qwen3, and gpt-oss on Hugging Face first, rather than training a custom speculator from scratch, to quickly validate inference cost savings.
DevOps & Infrastructure
Perplexity’s new agent runs entirely on your GPU — with one expensive catch
The New Stack
Perplexity's local agent, Portable Computer, is now available on Windows via download from the Microsoft Store, but it requires an NVIDIA GeForce RTX or RTX PRO GPU with at least 24GB of VRAM and is limited to Pro and Max subscribers. On Windows it runs Perplexity's own post-trained PPLX 27B model alongside Qwen 3.8 27B, both optimized for RTX GPUs, plus a built-in browser, tool calling, and Perplexity's proprietary SPACE sandbox. Connectors to Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub let the agent carry out real multistep workflow tasks. A key selling point is that work done locally draws no subscription credits and files never leave the device. The catch flagged in the piece is that the 24GB VRAM minimum shuts out most consumer GPUs, making the 'free' local compute effectively gated behind expensive hardware. The feature builds on Perplexity's existing Computer agent, now offloading execution to the user's own GPU instead of the cloud.
💡 Because this local-agent model trades cloud API costs for a hard 24GB+ VRAM GPU requirement, teams evaluating on-device AI agents need to weigh the API savings against the workstation-GPU procurement cost before assuming it's the cheaper option.
Digital Experience Monitoring with Grafana Cloud: Session Replay, synthetic checks, and faster investigations
Grafana
Grafana Cloud has tied Session Replay to synthetic checks to strengthen Digital Experience Monitoring (DEM). Synthetic Monitoring runs automated checks against critical user journeys to catch problems before real users hit them, while Session Replay in Frontend Observability visually replays what a user saw and did, correlated with Core Web Vitals, user actions, and traces. The core of this update is that every synthetic browser check run now automatically generates a matching Frontend Observability session. From a check's execution details page, a new 'View Frontend Session' button jumps straight into that run's session replay, user journey, and correlated traces. This lets engineers go from a failed synthetic check to determining real user impact without manually stitching together separate tools. Grafana says the combination can cut mean time to recovery (MTTR) from hours down to minutes.
💡 Auto-linking synthetic checks to session replays removes a manual triage step for confirming real user impact, making this a concrete observability upgrade worth adopting to shorten on-call responders' initial diagnosis time.
AI keeps finding security flaws — here’s what to fix first
The New Stack
The New Stack piece examines how AI-assisted scanning is finding vulnerabilities faster than security teams can triage them, turning discovery speed into an overwhelming backlog rather than a win. It opens with a security researcher testing a 300-person, globally-distributed B2B company who found an internet-exposed database with weak authentication rated critical severity — but the database turned out to be a resettable test instance used for job-candidate assessments, not a system holding client data. The example illustrates that CVSS severity alone can't set fix priorities; business context — whether an asset is internet-facing, what data it touches, and its role in revenue — is what turns raw findings into an actionable list. The recommended approach combines exploitability signals like CISA's Known Exploited Vulnerabilities (KEV) catalog and EPSS scores with attack-path analysis to cut thousands of alerts down to roughly 10-20 actionable tickets. The article also cites an academic study of more than 20,000 fixed issues finding that LLM-generated code fixes introduce nearly nine times as many new vulnerabilities as human-developer fixes, with distinct error patterns not seen in human code.
💡 As AI scanners inflate finding volume, prioritizing purely by CVSS severity stops working, so security and platform teams need a triage pipeline that combines KEV/EPSS exploitability signals and attack-path analysis with each asset's actual business context.
How Canvas Powers the AI Agent Development Feedback Loop
Honeycomb
Honeycomb's Canvas feature has reached general availability, and this post walks through how it supports every stage of the AI agent development feedback loop. The authors argue AI agents need a genuine feedback loop rather than purely reactive debugging: watch how the agent behaves, find what needs improvement, ship a change, and measure the result, ideally compounding into a flywheel over successive iterations. Canvas evaluates how well an agent's instrumentation adheres to OpenTelemetry's GenAI semantic conventions, flagging gaps or divergence, and checks whether the telemetry meets the requirements of Honeycomb features like Agent Timeline, the GenAI panel, and cost accounting. It also maintains a bank of common operational and business questions and checks whether your current telemetry can actually answer them. The post structures the loop into five stages: instrumenting agents with OpenTelemetry, understanding a single run, finding the problems worth fixing, shipping the fix, and proving it worked.
💡 With a GA tool now available to automatically check OTel GenAI semantic-convention compliance, teams should validate their agent instrumentation against the standard before production rollout, since it makes downstream cost and performance tracking far easier.
Datadog named the Company to Beat for observability platforms in 2026 Gartner® AI Vendor Race report
Datadog
Datadog was named the 'Company to Beat' for observability platforms in the August 2026 Gartner AI Vendor Race research, and separately says it was also named a Leader in the 2026 Gartner Magic Quadrant for Observability Platforms for a sixth consecutive year. The company describes a unified platform of more than 40 integrated products spanning observability, security, and AI-assisted operations, aimed at breaking down silos between IT ops, development, security, and business teams. AI-focused capabilities include Bits Investigation, which autonomously investigates alerts and recommends actions; Agent Observability, giving visibility into AI agent and LLM app performance, security, and cost; a Datadog MCP Server that lets AI agents query Datadog telemetry directly; Observability Pipelines with native OpenTelemetry support; and BYOC log management that keeps log data inside the customer's own cloud account. Recent moves include launching 100+ new capabilities at its DASH 2026 conference, acquiring Adaptive ML, and announcing a strategic partnership with Sakana AI. Looking ahead, Datadog names three priority challenges: managing costs as AI workloads generate large data volumes, data residency and sovereignty across regulatory jurisdictions, and governance and auditability of AI agents' execution privileges.
💡 Datadog's own list of priority challenges — AI data-volume cost, data sovereignty, and agent-execution governance — signals that teams evaluating its stack should make log-residency options (BYOC) and agent-permission auditability explicit requirements before signing a contract.
GitLab Dedicated: Compliance for a new regulatory era
GitLab
GitLab argues that regulations like the EU's NIS2 directive are no longer a distant planning consideration, citing the European Union Agency for Cybersecurity's (ENISA) NIS360 report as confirmation that supervisory authorities are now actively assessing cybersecurity maturity across critical sectors. The post positions GitLab Dedicated, its single-tenant managed offering, as a compliance vehicle, providing isolation, control, and audit-readiness to support regulations including DORA, NIS2, and GDPR. The exact granular features called out in the original post — such as specific regional data residency guarantees, audit-log retention periods, or named certifications — could not be confirmed in detail through available search results. Within what the excerpt and available material confirm, the post's apparent purpose is to position GitLab's single-tenant managed product as a compliance-driven offering aimed at organizations operating in Europe's tightening regulatory environment. The framing ties directly to NIS2's broader push for stricter supply-chain security and incident-reporting requirements across essential and important sectors.
💡 For organizations in scope of NIS2 or DORA, single-tenant managed hosting can be the fastest path to meeting isolation and audit-readiness requirements, so teams should weigh the migration cost to a Dedicated offering against the ongoing compliance burden of self-hosting.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.