Plumbing the
information age

The shopping agent lives inside apps or websites with integration points for catalog, cart, checkout, customer preferences, and order history. Customer requests like 'I need a tent, sleeping bag, and stove for a weekend trip with two kids' can trigger the agent to search the catalog and assemble multi-item requests. Anthropic also launched smart reports in beta for Claude Enterprise, which analyze team usage, costs, friction points, and identify repeated patterns worth packaging as reusable shared skills. Agent Skills and the Skills API (/v1/skills) came out of beta on the Claude API, with requests no longer requiring the skills beta header for Messages API calls that load Skills through the container parameter. This positions Claude Managed Agents as production-ready for e-commerce and retail use cases while simplifying the skill management layer for teams.

Nokia extended its partnership with Microsoft to accelerate network automation by developing an agentic, unified data foundation supporting faster, more reliable AI-driven operations for telecommunication providers. The collaboration integrates Nokia Data Suite's ready-to-use telco data products with Microsoft Fabric's unified analytics, governance and AI capabilities, enabling intelligent agent-based solutions across the network stack. Initial use cases focus on RAN optimization through autonomous VoNR assurance and geo-experience, which correlates subscriber, network and RF data. For practitioners, this represents a significant shift toward agentic automation in telecom networks—moving from manual data integration timelines spanning weeks to minutes-to-insight workflows. The unified data layer enables operators to deploy AI agents with trusted, governance-bound data access, reducing both integration toil and time-to-autonomy for network optimization tasks. The RAN optimization focus signals near-term deployment priority on carrier networks managing high complexity in 5G/5G-A rollouts.

SafeMind consists of two models: Red Tempest, an offensive model built to emulate AI adversaries, and Blue Solano, a defensive model designed to protect enterprise assets, connected by software harnesses that run them in a closed loop. CrowdStrike built these models using data from its Falcon sensors, which generate 7 trillion security events daily—proprietary data accumulated over 15 years that general AI companies cannot access. eCrime breakout time has shrunk to seconds, with average eCrime breakout time of 29 minutes and fastest recorded at 27 seconds. For security operations practitioners, this represents a critical shift: defenders now have purpose-built frontier models trained on CrowdStrike's proprietary 15-year security dataset, operating at machine speed to counter AI-driven threats that execute lateral movement in seconds rather than minutes.

Unified DDR consolidates previously fragmented detections into a centralized dashboard, enabling security teams to correlate identity, posture, and behavioral signals to surface critical incidents. This directly addresses alert fatigue in multicloud environments while extending protection to AI infrastructure—increasingly critical as organizations deploy autonomous agents and LLM workloads. Integration with Cortex Cloud's Data Security Posture Management (DSPM) capabilities and identity provider data enables contextualized incident enrichment, reducing noise and accelerating response decisions. For AIOps and SRE teams managing complex environments, this approach surfaces the most actionable incidents first while maintaining visibility across data planes that traditional SIEM cannot easily correlate.

Cornelis introduced Active Compute Fabric as an open architecture spanning scale-up and scale-out networking with programmable compute built into the fabric, marking the company's entry into scale-up networking and extending technology already powering AI and HPC workloads in hundreds of data centers. The design uses open industry standards including UALink and ESUN for scale-up and Ultra Ethernet specifications for scale-out, supporting multiple accelerator architectures. The economics are significant: GPU idle time represents approximately $1.68 billion annually in wasted capacity and 500 gigawatt-hours of power. Cornelis secured $205 million in funding, with CN5000 shipping today and CN6000 sampling with customers ahead of Q4 2026 availability. The company is collaborating with Qualcomm Technologies on rack-scale AI infrastructure, with Qualcomm's EVP joining Cornelis CEO Lisa Spelman at the September 15 AI Infra Summit keynote to discuss networking's role in AI system efficiency. For network practitioners, this represents a fundamental architectural shift: the network becomes an active compute participant offloading collective communication operations directly into fabric hardware rather than serving as passive transport.

At the AI Infra Summit 2026 in Santa Clara, Spelman revealed how Cornelis Networks is expanding beyond its established scale-out networking roots to address new challenges in AI infrastructure. The company announced the upcoming CN6000 second generation product and provided a first look at third generation technology, introducing a novel approach to network architecture. The conversation centered on how Cornelis is applying its heritage in low-latency, high-performance networking to emerging industry standards and new computing paradigms. CN5000 is currently shipping, CN6000 is sampling with customers, with broader availability expected in Q4 2026. The Active Compute Fabric design uses open standards (UALink, ESUN for scale-up; Ultra Ethernet for scale-out) to support multiple accelerator architectures. As rack-scale AI systems grow, both Cornelis and its partner Qualcomm see the network becoming a first-order design decision rather than an afterthought, shifting focus from traditional network provisioning to co-designed infrastructure where compute, memory, and networking are integrated from the start.

Axiado Corporation unveiled its approach to autonomous AI infrastructure management, addressing the critical constraint facing data center operators: power, cooling, security, and uptime have become the limiting factors, not GPU count. The company's Platform Efficiency Controller embeds intelligence directly into the management plane, enabling infrastructure to sense, decide, and act locally at machine speed. The Dynamic Thermal Management (DTM) Agent continuously adjusts thermal management based on actual platform conditions, delivering a 50% reduction in air cooling power. These reductions translate into 10–30% more tokens per dollar, allowing operators to generate greater useful AI output from existing infrastructure and power envelopes. Axiado CEO Gopi Sirineni presented "More Tokens Per Dollar: AI-Driven Platform Efficiency Controllers for High-Density AI Racks" at the conference on Wednesday, September 16. The announcement is directly relevant to platform engineers and infrastructure operators managing high-density AI deployments who face pressure to optimize both performance and operational costs in constrained power and cooling environments.

This guide addresses a critical confusion point in 2026 operations: the progression from MLOps (model lifecycle management) to LLMOps (large language model operations with prompt versioning and evaluation) to AgentOps (autonomous systems that use tools, make decisions, and execute multi-step tasks). MLOps made traditional ML manageable through CI/CD practices and drift monitoring. LLMOps added concerns around prompt versioning, retrieval, latency, cost per token, and hallucination detection. AgentOps introduces another operational layer: tracing multi-step reasoning chains, debugging tool call sequences, evaluating agent decision-making, and monitoring workflows where agents make dozens of LLM calls per request. The article clarifies that use cases determine framework choice: LLMOps applies to applications dependent on large language models, retrieval, and generated responses; AgentOps applies when systems include agents that use tools, make decisions, and complete multi-step tasks autonomously. For SREs and platform engineers deploying agentic systems, understanding these boundaries is essential for correct observability architecture, as traditional APM and monitoring tools designed for deterministic software fail to capture agent reasoning and tool execution traces.

Monid addresses the operational bottleneck of agent tool discovery and cost optimization by providing a unified interface for agents to access endpoints across multiple categories: SEO, leads, social, search, ecommerce, stocks, media generation, 3D modeling, and private company data. The tool implements a freemium model where discovery and inspection are free, with pay-per-call pricing, eliminating subscription overhead for infrequent tool use. The stack supports Deno 2 self-hosting and declarative TypeScript connectors, allowing both hosted and forked deployments. The 100% month-over-month transaction growth (4M to 8M in two weeks) suggests significant adoption among developers building multi-agent systems. For platform engineers evaluating agent infrastructure, Monid demonstrates a critical operating pattern: agents need lightweight, dynamic tool binding without the overhead of traditional integration platforms. The open-source design with fork-friendly architecture allows internal deployments, addressing data residency and governance concerns common in enterprise agent adoption.

Anthropic significantly scaled Claude for Small Business from initial connectors to 43 pre-built workflows and 27 deep integrations spanning QuickBooks, PayPal, HubSpot, Canva, Docusign, Google Workspace, Microsoft 365, plus new integrations with Shopify, Salesforce, TikTok, Atlassian, Zoom, Xero, Gusto, Square, Stripe, and Zapier. Over 900,000 organizations have adopted the plugin since May. The update addresses core SMB pain points in sales conversion, financial management, and reporting. Anthropic is backing the launch with instructor programs: free in-person workshops across 10 U.S. cities and a 14-partner webinar series running late September through November 2026, with Approved Claude SMB Trainers conducting over 750 community workshops. This represents real agentic workflow automation with measurable adoption, not marketing positioning.

The DeepMind Institute represents a formal institutional response to accelerating AGI timelines. Co-directed by Shane Legg (Chief AGI Scientist), Demis Hassabis (Chair, Alphabet Chief Scientist), and James Manyika (Google SVP Research), the platform brings together Google, DeepMind, and external researchers to publish peer-reviewed perspectives on AGI governance, safety, and economics. Inaugural research addresses Universal Basic Capital, reasoning model transparency, catastrophic biosecurity safeguards, and international governance. The institute explicitly acknowledges contributors will not always agree and may revise positions as frontier evidence changes. Legg reaffirmed his 50% probability of minimal AGI by 2028 and stated Anthropic CEO Dario Amodei's call to slow releases is "worth considering." This signals major labs moving beyond isolated safety research into formal governance and economic infrastructure—an intellectual positioning that extends beyond product announcements.

Nokia partnered with Microsoft to develop an agentic, unified data foundation supporting faster, more reliable AI-driven operations for telecoms providers. The technical foundation combines Nokia's established telco data products with Microsoft's enterprise analytics platform to unify disparate operational data sources. The offering combines Nokia Data Suite's telco products with Microsoft Fabric's unified analytics, governance and AI capabilities, enabling intelligent, agent-based solutions across the network stack. For network operations practitioners, this addresses a longstanding pain point: data silos across OSS/BSS, RAN, and core network systems fragment the context available to autonomous agents. By collapsing latency from weeks to minutes, operators can deploy agentic workflows with current, trusted telemetry—critical for autonomous fault management, performance optimization, and service provisioning. Agent decisions are only as good as their input data; stale or inconsistent information undermines autonomous operations at scale.

The first phase of the AI infrastructure boom has created enormous demand for compute, power and optical connectivity inside data centers, with hyperscalers needing high-capacity fiber connecting data centers across metropolitan and long-haul routes. Verizon's strategy extends beyond traditional data center interconnect: as AI moves from centralized training toward inference, Verizon is repurposing thousands of legacy central offices as locations for edge compute. Verizon expects AI Connect construction to use success-based capital, with investment tied to specific opportunities rather than speculative network expansion. For network operators, this represents a tactical shift: rather than speculative fiber builds, Verizon is matching capital deployment to signed customer contracts, reducing risk exposure during demand uncertainty. Central offices already hold power infrastructure, fiber aggregation points, and real estate—repurposing them for inference workloads avoids greenfield data center construction while leveraging existing assets. Powered, permitted and fiber-connected central offices put inference infrastructure closer to enterprises, users and physical AI applications.

Trump's statement directly contradicts the industry safety push announced by Amodei, Altman, and Musk. The response signals fundamental policy misalignment between frontier labs' internal governance proposals and the administration's national security lens, which views AI development speed as core to U.S. strategic competition with China. While Trump suggested guardrails could be implemented, his framing prioritizes geopolitical dominance over safety coordination, creating immediate pressure on companies seeking both internal safety measures and federal alignment on AI policy.

This article frames the structural tension blocking effective governance: safety pacing works only if all major actors participate simultaneously, but no state has credible verification that others will comply. The piece complements both the industry safety movement and Trump's security-first positioning by examining why individual actor rationality (accelerate to avoid falling behind) leads to collectively irrational outcomes. Neither voluntary corporate coordination nor international treaties appear likely to overcome this dilemma in 2026, making enterprise compliance frameworks like the EU's enforcement wave the only enforceable lever available.