Plumbing the
information age

ServiceNow's Q2 results validate its AI platform strategy by tying AI adoption to governed action rather than assistant-only use cases. AI Control Tower, built on the Traceloop acquisition completed in March 2026, provides deep runtime observability into how AI agents reason, where they make decisions, and when course-correction is needed. This replaces periodic manual audits with continuous live monitoring. The platform now discovers AI assets across 30 third-party systems including AWS, Google Cloud, Microsoft Azure, SAP, Oracle, and Workday—extending governance, observability, and security well beyond ServiceNow's own ecosystem. For network and AIOps practitioners, this signals a fundamental shift: observability platforms must now monitor not just infrastructure and applications, but the behavior and decision-making of autonomous AI agents. The practical implication is that traditional alert-and-dashboard models are giving way to runtime governance and tracing, requiring new instrumentation and decision-tracking capabilities at the agent execution layer.

On July 28, MCP's maintainers will finalize the protocol's 2026-07-28 revision with the most substantial changes since adding authorization, including removal of the initialization handshake. The protocol version, client info, and client capabilities now travel in _meta on every request. The 10-week window between the release candidate and final spec allows SDK maintainers and server implementers to validate changes against real workloads—critical for teams already running MCP servers in production. For infrastructure practitioners, the shift to statelessness means MCP servers can now be horizontally scaled behind load balancers and API gateways without session affinity concerns, enabling MCP as enterprise middleware rather than point solutions. Stacklok CEO Craig McLuckie noted that MCP's original stateful design emerged from supporting developers using coding tools locally, but the broader AI community now embraces MCP as a mechanism for enterprises to intermediate access to production systems with consistent governance and operational controls.

The MCP 2026-07-28 specification is not a marginal update but a fundamental architectural shift enabling MCP as critical infrastructure behind load balancers, API gateways, and enterprise identity providers. Release candidate locked May 21, 2026; final spec ships July 28, 2026. This follows April 2026 analysis of how MCP stabilization and A2A protocol reshaped agent interoperability. The move to stateless design is operationally critical: MCP servers can now be deployed as stateless microservices on Kubernetes, scaled horizontally with standard load-balancing, and integrated into existing service mesh ecosystems without session affinity. Authentication and session state move to the HTTP layer (OAuth/OIDC), enabling enterprises to apply existing identity and access management infrastructure. For companies already connecting Claude, Cursor, n8n, or custom agents to remote MCP servers, this clarifies migration paths: what breaks, what improves, and how to upgrade without interrupting workflows. The stateless design also reduces operational complexity—MCP becomes just another HTTP service in your infrastructure, governed by standard observability, scaling, and resilience patterns.

Cisco's Secure Network will underpin the PCs, which are designed for powering intense AI and agentic workloads from the desktop. For companies deploying Halo PC en masse for their developers, Cisco will support fleet-wide observability, tracking agent behavior, token usage, and compute utilization. DefenseClaw, the vendor's open source agent security tool, adds guardrails directly on-device, ensuring policy enforcement is maintained. For infrastructure teams scaling agentic AI at the edge, this represents a practitioner-relevant convergence: observability tools built for cloud-native environments now extend governance, policy enforcement, and telemetry collection to distributed AI developer endpoints. The partnership surfaces the operational reality that autonomous AI agents deployed across infrastructure require centralized visibility and control—whether running in data centers or developer workstations—and shows Cisco positioning observability as the foundation for managing agentic systems at scale.

CrowdStrike and Cerebras announced a partnership to integrate Cerebras' AI inference infrastructure with CrowdStrike's Falcon AIDR models, with Cerebras also adopting Falcon to secure its own operations. The timing addresses a critical shift in the threat landscape where every millisecond determines whether AI prevents an attack or merely explains it afterward. For SecOps and AIOps practitioners, this partnership has direct operational implications. CrowdStrike will leverage Cerebras's industry-leading inference speed to run Falcon AIDR models at machine speed, pairing AI-native security with the world's fastest inference. The pairing of Falcon's lightweight agent architecture with Cerebras' inference capability targets the detection-to-response latency gap that manual SOAR playbooks cannot close. Shares of Cerebras surged 7.05% while CrowdStrike slipped 1.90% on the announcement. For organizations standardized on Falcon, this partnership signals deeper investment in AI-native detection. For those evaluating AI-augmented incident response, the inference-speed component becomes a measurable procurement factor when comparing XDR and AIDR platforms.

AMD formally launched its 6th generation EPYC server CPUs at its flagship Advancing AI conference July 22-23, 2026. Venice represents a major generational leap: it's the first x86 HPC CPU to achieve volume production on TSMC's 2nm node, with up to 256 cores (using compact Zen 6c) and up to 96 full-speed Zen 6 cores depending on SKU. The processor features 5th generation Infinity Fabric, PCIe Gen 6 with 128 lanes per socket, 16-channel DDR5 memory support (up to DDR5-12800 on MRDIMMs), and CXL 3.1 support. For networking practitioners, the CPU includes significantly increased I/O die capacity to support four MI450-series GPUs, DPUs, and 800G NICs per node in Helios racks. The Helios rack-scale system bundles 72 MI455X GPUs with Venice CPUs and Pensando networking into a single integrated platform priced around $5.25M, with H2 2026 availability. This creates a real procurement window for enterprises: with Intel Xeon's next competitive response not arriving until mid-2027, AMD's integrated stack (CPUs, GPUs, networking, orchestration) now represents the only current-generation alternative to hyperscaler-optimized systems.

Arista Networks continues to dominate the data center Ethernet switch market with strong momentum from AI infrastructure buildouts. The company reported Q1 2026 revenue of $2.71B (35% YoY growth), with deferred revenue reaching $6.2B—driven by longer customer acceptance cycles on new AI leaf-spine products rather than demand weakness. The company raised its full-year AI networking revenue target to $3.5B, effectively doubling YoY AI sales. Two customers (Microsoft and Meta) each represent 10%+ of total revenue, underscoring concentration among hyperscalers. The technical transition reflects: shift from 400G→800G→1.6T port speeds; power density as a primary constraint alongside bandwidth; adoption of liquid cooling and linear-drive pluggable optics (LPO) to reduce power in dense AI fabrics; and rapid adoption of Broadcom Tomahawk 6 silicon in next-generation platforms. Arista's 7060XE7 series (announced June 9, 2026) targets rack-scale AI with 1.6T ports, shipping Q4 2026. Meanwhile, Nvidia has captured 21.5% market share in data center Ethernet switches (Q1 2026), creating competitive pressure on all incumbent vendors.

Claude Opus 5 first surfaced as "Honeycomb EAP" inside Cursor on July 9 and was pulled within hours, then rolled out across providers on July 23, 2026, with the rollout extending into July 24. Standard mode is $5 per million input tokens and $25 per million output tokens — the same rate as Opus 4.8 and half of Fable 5's input price, while fast mode doubles the rate ($10/$50) and runs about 2.5× faster. The spec profile includes a 1 million token context window, an "xhigh" reasoning effort setting, per-turn controls, and a safety fallback that routes to Claude Opus 4.8 on trigger. The model outperformed Opus 4.8 across all internal life sciences benchmarks and proved particularly adept at automated protein research, an important field that was the focus of the 2024 Nobel Prize in chemistry. On Anthropic's own table it more than doubles Opus 4.8 on agentic terminal coding and posts the lowest misalignment score the company has measured. Anthropic discloses that its Frontier-Bench v0.1 figures come from an internal run on the mini-SWE-agent harness with a GKE backend, taking mean reward over five attempts per task — and that Opus 4.8 served as the fallback on safety-classifier refusals.

Building on Gemini 3.5 Flash, Google describes 3.6 Flash as its workhorse model, reducing output token usage by 17% according to the Artificial Analysis Index, and released three new Gemini models on July 21: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. Output dropped from $9.00 to $7.50 per million tokens with input held steady at $1.50, and on the DeepSWE coding benchmark the model scores 49 percent, up from 37 percent for 3.5 Flash. Gemini 3.6 Flash scores 49% on the DeepSWE benchmark and pushes machine learning engineering performance higher, scoring 63.9% on MLE-Bench compared to 49.7% previously. The model takes fewer reasoning steps and tool calls to accomplish multi-step workflows and exhibits reduced verbosity. Both 3.6 Flash and 3.5 Flash-Lite feature a one-million-token input context window and a maximum output limit of 64,000 tokens, with native multimodal input, thinking controls, and built-in Computer Use as a client-side tool via the Gemini API.

The challenge for enterprises is making AI agents reliable enough to do high-value work in production, which requires more than a model: it requires systems, evaluations, and deployment expertise to improve agents without giving up control. Deployments are led by OpenAI Forward Deployed Engineers and select global systems integrators, drawing from the OpenAI Deployment Company subsidiary formed in May with $4 billion in backing and a $10 billion valuation. Target use cases span customer support, outbound sales development, procurement, IT services, and HR functions, with agents able to verify identity, access account data, apply company policies, and execute approved actions including billing resolution and refunds. OpenAI says Presence already handles its own English-language phone support line and resolves 75% of inbound calls without human intervention, with a Codex-driven improvement process reducing human handoffs by 15 percentage points within 10 days of launch. Presence is not available as a self-service product, so interested companies must work through their OpenAI account team.

The acquisition consolidates a critical vertical AI platform into Dassault's life-sciences product suite. ArisGlobal specializes in FDA and regulatory-compliance workflows for pharmaceutical and medical device manufacturers, processing 12+ million patient-safety reports annually and serving approximately 50% of the world's top 50 biopharma companies. The deal structure—$1.8B upfront plus $200M in multi-year AI revenue milestones—reflects strategic intent to embed AI-native regulatory capabilities across Dassault's installed base rather than treat it as a standalone acquisition. Expected close is H2 2026 pending regulatory approval. For enterprise technology leaders, this acquisition exemplifies the 2026 M&A pattern: buyers with broad installed bases acquiring deep vertical AI specialists to accelerate time-to-value and avoid building domain-specific AI capabilities from scratch. The milestone structure also signals that acquirers now value demonstrated revenue traction in AI-augmented workflows, not just technical capability. Organizations evaluating AI compliance tools should anticipate consolidation among standalone vendors and rapid integration into larger enterprise stacks, which may accelerate feature deployment but increase switching costs.