🔥 Top Story
Cloud Network Insights: end-to-end observability for the Cross-Cloud Network
Google Cloud unveiled Cloud Network Insights, providing end-to-end observability for the Cross-Cloud Network. As networks span beyond a single data center or cloud, the service continuously monitors agent, network, and web performance across Google Cloud, AWS, Azure, data centers, internet applications, and agentic workloads. Built as a first-party Google Cloud offering that integrates Broadcom's network observability (AppNeta), it extends visibility to paths you don't own, such as ISP links and third-party clouds. Its core technique is active synthetic probing, which exercises network routes even when no user traffic is present so issues can be caught before they hit users. It enables rapid root-cause analysis, distinguishing network problems from application- or browser-level issues, and aims to cut mean time to resolution (MTTR). It is delivered as part of Network Intelligence Center.
💡 Why it matters: Being able to trace segments outside your own network in one place is practically valuable for ops teams that must quickly assign blame across cloud, ISP, and app when something breaks.
🔗 Read more · Google Cloud
Kubernetes & Cloud Native
Why cloud native belongs at the heart of agentic AI: Lessons from building a multi-agent security platform on Kubernetes
CNCF
Willem Berroubache, Lead Security Architect at Orange Innovation and a CNCF Golden Kubestronaut, distills from his KubeCon + CloudNativeCon Europe 2026 (Amsterdam) talk why agentic AI belongs on cloud native. The case study is a multi-agent security platform (SOC automation) Orange Innovation is building and rolling out, with Detect (ML), Analyse (threat analyst), Remediate, and Notify (Mattermost) agents plus a Human-in-the-Loop branch and a feedback loop that retrains the anomaly model. For cost and accuracy, a Kafka stream is pre-filtered by a classical Isolation Forest anomaly model before anything reaches the LLM-driven agents. Each agent is deployed as its own Kubernetes Deployment with its own resource limits, identity, and restart policy, and communication is controlled with CiliumNetworkPolicy rather than a service mesh. Inter-agent coordination uses the A2A protocol (open-sourced in 2025, under Linux Foundation governance) and environment integration uses MCP. The standout lesson: encode the reviewer agent's safety constraints not in a system prompt but as OPA policies and Kyverno admission rules, calling OPA via MCP to get a deterministic verdict.
💡 Externalizing an agent's safety constraints into version-controlled, testable policy (OPA/Kyverno) instead of prompts, and pre-filtering LLM calls with classical ML, is a practical blueprint for getting both reliability and cost control in security automation.
AI & ML
MolmoMotion: Language-guided 3D motion forecasting
Hugging Face
The Allen Institute for AI (Ai2) released MolmoMotion, a language-guided 3D motion forecasting model. It takes an RGB observation, a set of query points on an object, and a text description of the intended action, and predicts the object's future 3D point trajectory. The model uses Molmo 2 as its backbone to connect language instructions to objects and points in an image. It comes in two variants: the autoregressive MolmoMotion-AR writes future 3D coordinates as structured text step by step (strongest when the path is well-defined), while the flow-matching MolmoMotion-FM transforms noise into motion in continuous 3D space, better capturing uncertainty when multiple futures are plausible. Ai2 is openly releasing the model weights, the MolmoMotion-1M dataset, and the PointMotionBench benchmark; the dataset spans 736 motion types and was built by an automatic pipeline that extracts object-grounded 3D trajectories from unconstrained video. The predicted trajectories can directly drive downstream tasks such as robot planning and trajectory-conditioned video generation.
💡 By forecasting how things will move next as open 3D point trajectories rather than just observing motion, it offers a lightweight motion representation that plugs straight into both robot policies and video generation.
From the Hugging Face Hub to robot hardware with Strands Agents and LeRobot
Hugging Face
Amazon demonstrated a workflow from a Hugging Face Hub dataset all the way to physical robot hardware, via an integration of its Strands agent framework with LeRobot. The key idea is that the Strands Robots SDK exposes the LeRobot stack as AgentTools, composing data recording, simulation, and policy execution into a single Strands agent loop. Because simulation and hardware share the same on-disk LeRobotDataset format, data recorded in sim and on a real robot is handled identically, and policies are swapped with a single string (keyword argument). For multiple robots, a built-in peer mesh lets the agent coordinate a whole fleet. The default path runs entirely in laptop simulation (MuJoCo backend, Linux/macOS including Apple Silicon) with no hardware, GPU, or Hugging Face credentials required, and you can clone a sample app from GitHub to run it. Later steps swap in trained policy checkpoints such as GR00T or MolmoAct2 to extend to real grasping on SO-100/SO-101 robots.
💡 Sharing one data format and agent code across sim and hardware, and swapping policies with a single argument, sharply tightens robotics' sim-to-real loop and lowers the hardware/GPU barrier to entry.
A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry
OpenAI
OpenAI, with drug-discovery startup Molecule.one, showed how a near-autonomous AI chemist improved a challenging reaction in medicinal chemistry. OpenAI's GPT-5 reviewed the scientific literature, generated and ranked research proposals, designed experiments, analyzed results, and proposed follow-ups. Human chemists steered the work, chose which proposals to test, and validated the final result, a near-autonomous collaboration rather than full automation. The target was a widely used drug-discovery reaction whose variant involving primary sulfonamides historically gave low yields, limiting its usefulness; the model proposed an unexpected way to improve it. Paired with Molecule.one's Maria AI and a dedicated lab, the proposals were physically validated, and the full process took about 2.5 months, plus roughly half a month to write up results. The companies frame it as an early case of AI contributing to an open-ended scientific problem in organic chemistry.
💡 With AI assisting every stage of the research cycle, from literature review to experiment design, analysis, and validation under human supervision, it shows real scientific output hinges less on raw model power than on the human-AI division of labor wired into lab automation.
GLM-5.2: Built for Long-Horizon Tasks
Hugging Face
Z.ai released GLM-5.2, a large language model built for long-horizon tasks. It claims a substantial leap in long-horizon capability over its predecessor GLM-5, and rather than merely advertising a 1M-token context, it substantially expanded 1M-context training for coding-agent scenarios such as large-scale implementation, automated research, performance optimization, and complex debugging. A "Max effort" level lets users allocate extra computation to harder tasks to push coding performance further. Technically, IndexShare, multi-token-prediction (MTP) layers, and KV-cache FP8 ease the long-context bottleneck (which shifts from compute to KV-cache capacity) and reduce training-inference mismatch. For reinforcement learning, the team moved from group-wise comparison to critic-based PPO that learns token-level advantages from individual rollouts, with safeguards against reward hacking, for example agents reading protected evaluation artifacts or answers. All of this runs on Z.ai's own "slime" infrastructure spanning training to large-scale inference rollout, with OPD training merging more than ten expert models in about two days for efficiency.
💡 It shows that actually sustaining 1M context and long-horizon agent tasks requires optimizing not just model quality but the KV cache, inference engine, and RL infrastructure together, a strong signal for teams weighing self-hosted coding agents.
Introducing LifeSciBench
OpenAI
OpenAI introduced LifeSciBench, a benchmark for evaluating how well AI systems handle real-world life-science research tasks and decisions. It comprises 750 expert-authored, expert-reviewed tasks spanning seven workflows: evidence handling, analysis, design and optimization, scientific reasoning, validation and operations, translation, and scientific communication. It was developed with 173 scientists from biotech and pharma, and each task pairs a scientific prompt and relevant context or artifacts with a free-response answer graded by expert-written rubrics. About 79% of tasks require multiple reasoning or decision steps, an average of four, so they reflect real research flows rather than simple factual lookups. The benchmark tests whether models can reason from evidence, work with scientific artifacts, handle uncertainty, and make useful decisions under real-world constraints.
💡 By grading models on multi-step research workflows with expert rubrics rather than short-answer accuracy, it offers a template for evaluation design that gauges whether domain-specific agents are actually fit for real work.
Agentic Resource Discovery: Let agents search
Hugging Face
Agentic Resource Discovery (ARD), an open draft specification pushed by multiple companies, was unveiled. Agents today use a handful of protocols like MCP and A2A, but all of them assume the user already knows which tool or agent they need. ARD is the discovery layer that sits in front of them, defining how agents and tools are cataloged, indexed, and searched across federated registries so an agent can find capabilities at runtime instead of needing them pre-installed. In other words, it is a shift from an install-first model to intent-based search. The spec is an open standard with contributors from Microsoft, Google, GoDaddy, Hugging Face and others, and any company can implement it independently. Hugging Face implemented it as "Discover," serving the Hub's semantic search over Spaces and its Agent Skills as ARD catalog entries, searchable via a REST API or an MCP server.
💡 If a standard for discovering and invoking tools at runtime, rather than pre-installing and configuring them, takes hold, agents could reach thousands of ad-hoc tools and services on demand, addressing the real scalability gap in the MCP/A2A ecosystem.
Cloud Updates
How growing UK midsize businesses are building in the AI era
Google Cloud
Timed to London Summit 2026, Google Cloud highlighted how growing UK small and midsize businesses (SMBs) are building in the AI era. It notes that the UK's 5-million-plus SMBs are the backbone of the economy and are now putting AI to work to operate more efficiently, move faster, and deliver better outcomes for customers. The piece showcases the motivations and examples behind this SMB AI adoption, stressing that cloud and AI are a competitiveness lever not just for large enterprises but at smaller scale too. Google Cloud ties the trend to its own platform, positioning itself to help SMBs adopt AI without high barriers. Overall it is a market and customer-perspective article about the UK SMB AI transition rather than a product launch.
💡 Accelerating SMB AI adoption means a new base of demand for cloud providers, and from an engineering view it suggests standardized managed AI services and low operational burden are the key adoption variables.
From AI potential to agentic reality: Driving the UK’s next chapter
Google Cloud
Marking London Summit 2026, Google Cloud framed the UK as driving the next chapter from AI potential to agentic reality. The piece assumes the UK, and London in particular, remains one of the great hubs for AI development in Europe and the world. Its core message is that enterprises are moving beyond experimental generative AI toward agentic AI that actually performs work. Google Cloud connects the infrastructure, data cloud, and agent-platform capabilities that support this shift to the UK market. Overall it is a vision-led article emphasizing UK AI leadership and enterprise agent adoption rather than specific product specs.
💡 The narrative of shifting from generative demos to agents that do work also implies the practical challenge that agent adoption stalls unless infrastructure, data pipelines, and governance keep up.
Build and Deploy a Remote MCP Server to GKE in 30 Minutes
Google Cloud
Google Cloud published a hands-on tutorial for building and deploying a remote MCP (Model Context Protocol) server to GKE (Google Kubernetes Engine) in about 30 minutes. The motivation is that integrating context from tools and data sources into LLMs is challenging and raises the difficulty of developing AI agents. MCP is a protocol that standardizes this tool-and-data integration, and the tutorial walks step by step through containerizing a remote (rather than local) MCP server and running it on a GKE cluster. This lets multiple agents and clients reach a shared MCP server over the network and use tools and context consistently. Deploying on Kubernetes brings cloud-native benefits, namely scalability, availability, and operational standardization, to the MCP server.
💡 Running an MCP server as a remote GKE service instead of a local process centralizes tool access across teams and agents, letting you scale, observe, and access-control it with standard Kubernetes practices.
DevOps & Infrastructure
“A data lake of nuance for AI agents to swim in”: AWS Context gets shipshape on reasoning
The New Stack
AWS launched Context, a knowledge graph service that gives AI agents governed access to enterprise data relationships, rules, and domain knowledge. The piece starts from the observation that simply feeding agents ever more data ("all-you-can-eat") doesn't improve their reasoning. Rather than serving raw data, Context structures the relationships between data, business rules, and domain knowledge so agents can reason over meaning and constraints, not just rows. The goal is sharper, more grounded agent reasoning via a connected graph instead of an undifferentiated data pile. Governance is central: the service emphasizes controlling which agents can reach which relationships and knowledge. It reflects a broader push to standardize an enterprise semantic layer beyond plain RAG or vector search.
💡 It reframes agent reliability around context with relationships and rules rather than sheer data volume, making internal data governance and semantic-layer design prerequisites for trustworthy agents.
Kiro goes mobile: AWS brings agentic coding supervision to the iPhone
The New Stack
AWS released a native iOS app for Kiro, its AI-powered development environment. The core premise is that the agent keeps working even after you walk away from your desk. The new mobile app lets developers monitor in-progress agentic coding sessions, steer them, and approve work from their phones. It targets a workflow for supervising long-running agent tasks asynchronously, untethered from the desktop. In effect, it extends human-in-the-loop oversight, where agents generate code autonomously but humans intervene and approve, onto mobile. The significance is keeping the supervision point available even when you are away from your desk.
💡 As agentic coding runs long and asynchronously, the supervision point is moving off the desktop, and a mobile approval flow hints at a new operating pattern for developers managing parallel agents.
Vercel launches eve, an open-source framework that treats agents as directories
The New Stack
Vercel on Wednesday launched eve, an open-source framework for building AI agents that treats each agent as a directory. The central design idea is to represent an agent's code, configuration, and resources as a filesystem directory rather than a complex runtime abstraction. That gives developers a familiar file-tree mental model and makes agent configurations easy to version, port, and clone. Vercel released eve as open source so the community can extend and contribute to it. Modeling agents as directories points toward treating agents like code. The approach aims to fold agent development into familiar, file-based workflows.
💡 Standardizing agents as directories/file-trees lets teams layer Git-based versioning, code review, and CI directly onto them, making agent operations easier to fold into existing software delivery pipelines.
How Block manages its fleet of AI coding agents from Slack
The New Stack
Fintech company Block detailed how it runs a fleet of AI coding agents across hundreds of services. Most AI coding tools work well inside a single repository, but few can operate across many services, which is exactly what Block needed. To solve this, Block built its own agent orchestration system on top of Goose, its open-source agent. The control surface is Slack: developers dispatch work to the agent fleet and get results in chat, a ChatOps-style workflow. The significance is a real-world pattern for scaling and coordinating agents across an entire organization's codebase, not just one repo.
💡 Pulling agents out of the single-repo IDE and running them as a Slack-based fleet signals that org-wide code changes are becoming a platform-team orchestration problem rather than something humans launch one by one.
Chainguard Agent Skills matures
The New Stack
Chainguard, a container and software-supply-chain security company, is expanding its effort to secure AI coding agents. Responding to the fast-growing agent ecosystem, it introduced a new public registry for the skills that agents use. The underlying concern is that unvetted skills and tools pulled in by agents become a fresh supply-chain attack surface. Chainguard brings its core strength, delivering trusted and verified artifacts, to agent skills, so skills with assured provenance and integrity can be published and consumed. It is an attempt to standardize software-supply-chain security for the agent era.
💡 When agents dynamically pull in external skills, the skills themselves become a supply-chain threat surface, so a registry of verified skills could become as basic a control as signing container images.
AWS puts an AI bouncer at the merge queue
The New Stack
AWS is pushing its DevOps Agent into the delivery pipeline itself, putting an AI bouncer at the merge queue. The premise is that as AI floods repos with generated code, the bottleneck has shifted from writing code to verifying and releasing it. The new capabilities, release readiness review and autonomous release testing, automatically check whether changes are fit to ship at the front of the merge queue. In other words, agents don't just produce code; another agent gatekeeps that code's release-worthiness. It is an attempt to preserve quality and stability when changes arrive faster than humans can review them one by one.
💡 As code generation gets cheap, the bottleneck moves to verification and release, so an automated gate on the merge queue becomes an essential control for safely absorbing change volume beyond human review capacity.
“Agents need boring infrastructure around them”: Why we need to take an interest in ‘invisible’ AI
The New Stack
Tailscale expanded its Aperture product while arguing that agents need boring infrastructure around them. The concern is that AI is already inside most enterprise IT stacks, but employees use it in a shambolic, unsystematic way that creates blind spots. The update adds a chat interface, MCP/API connectors, and sandboxes to Aperture. The core is identity-based control: enterprises can govern which agents and users can reach which LLMs, tools, and data on a per-identity basis. The message is that the prerequisite for running agents isn't a flashy model but invisible infrastructure, namely access control, isolation, and connectivity.
💡 The real risk in adopting agents is not model quality but the absence of control over who can access what, and identity-based access control plus sandboxing is the starting point for pulling shadow AI into a manageable surface.
Google, Microsoft, and OpenAI join forces to help create AI’s missing trust layer
The New Stack
Google, Microsoft, OpenAI and others joined forces to build AI's missing trust layer, launching a new body called the Appia Foundation. With 13 founding members, the foundation is a Linux Foundation project aimed at making AI safety claims verifiable. The concern is that vendors assert their AI is safe but there's no standardized way for third parties to confirm it. Appia wants to establish a common framework and verification mechanisms for safety claims, turning marketing-style assertions into objectively demonstrable evidence. The Linux Foundation participates as a neutral governance host, pushing for standards not tied to any single vendor.
💡 A neutral standard that converts it-is-safe assertions into verifiable evidence could give enterprises a concrete basis for evaluating AI systems during regulation and procurement.
Your AI isn’t broken. Your data is.
The New Stack
Under the provocative headline "Your AI isn't broken. Your data is.", a startup called Clario introduced an unstructured-data cleanup platform. It pins the reason enterprises pour billions into AI yet get garbage back on data quality. The target is so-called ROT files, redundant, obsolete, and trivial data, the noise that silently sabotages enterprise AI projects. Clario's platform identifies and removes this ROT to raise the signal-to-noise ratio of the data used for training and retrieval. Its pricing is unusual: a results-based model that only gets paid once the garbage is gone.
💡 The point that many RAG and fine-tuning performance problems stem from ROT data rather than the model is a reminder that the first step of an AI project should be data cleanup and governance, not model selection.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.