Plumbing the
information age

The roadmap, developed by Core Maintainers and Working Groups, identifies five priority areas organized around production deployment needs. Following the 2026-07-28 release that transformed MCP from bidirectional stateful to request/response stateless, the protocol now supports deployment like any other HTTP workload. Tool calling—the most-used MCP feature—is being improved through standardized result handling and better scaling for large tool surfaces. The roadmap addresses a key pain point: models pay token costs upfront for all tool definitions before users ask questions, and tool selection degrades with list size. MCP crossed 400 million monthly SDK downloads in 2026, a 4x increase, signaling production-scale adoption. This maintenance-focused update reflects the shift from protocol novelty to operational reality for enterprise agent infrastructure.

Anthropic's flagship Claude Opus 5 arrived July 24, 2026 with unchanged pricing and was updated August 12 with improved inference speed and scientific research capabilities, becoming the new default for Claude Max with near-Fable-5 performance at half the cost, a 5-level effort toggle for dynamic reasoning depth, 1M-token context plus 128K max output, and knowledge cutoff of May 2026. This defining opportunity of 2026 for enterprises struggling with speed-to-value shows agents are no longer experimental—they're production infrastructure. August 2026 created a fundamental challenge with benchmark lag: models ship faster than independent verification can complete. For SRE and AIOps practitioners, this acceleration means the agent harness landscape is fragmenting (Claude Code, Codex CLI, Muse Code, Gemini, Grok), making tool standardization through MCP more critical for avoiding vendor lock-in.

CrowdStrike's Fal.Con 2026 sold out faster than any previous year, bringing together 10,000+ attendees from 4,000 organizations across 71 countries, backed by a record 150 ecosystem sponsors led by AWS, Accenture, Anthropic, Dell, EY, Google Cloud, Horizon3, Intel, NVIDIA, and OpenAI. What makes this noteworthy for operations practitioners is not just the scale, but the composition: the presence of foundational AI companies (Anthropic, OpenAI, Google Cloud), infrastructure providers (AWS), and implementation partners (Accenture, EY) signals that 2026 is the year enterprise AI-powered security operations move from pilot to production. The surge in demand reflects a new reality: the more AI organizations adopt, the more cybersecurity they require. For network and SRE teams, this confirms that agentic SOC automation—not just tooling, but orchestrated, autonomous response workflows—is now table stakes. The conference emphasis (Aug 31–Sep 3) comes right before CrowdStrike's broader Fal.Con announcements on product innovation, likely to include expanded Falcon platform integrations that practitioners will need to evaluate for MDR, SIEM consolidation, and automated incident response workflows.

Intezer's new Workflows product is a native automation and response builder that enables security teams to create and customize response workflows directly inside the platform, consolidating triage, investigation, and response without requiring a separate SOAR system. For AIOps and security operations practitioners, this represents the broader industry trend of tightening the feedback loop between detection and response—what Intezer calls "bringing response into the same platform where alerts are triaged." The motivation is clear: enterprises managing thousands of daily alerts using separate tools (XDR → SIEM → SOAR) accumulate context-switching overhead and increase mean time to respond (MTTR). By embedding native workflow automation inside the investigation platform, Intezer reduces tool sprawl and accelerates response by eliminating cross-platform data handoffs. This also parallels what Fortinet announced earlier this year with FortiSOC (unifying SIEM, SOAR, and threat intelligence) and what CrowdStrike is positioning with agentic SOC capabilities. For practitioners evaluating consolidation strategies, this is a data point: vendors across the security stack are racing to embed automation closer to detection and investigation logic, not as optional add-ons.

MetaRoCE represents a clean-sheet redesign of RDMA transport for hyperscale infrastructure. Rather than assuming lossless Ethernet like traditional RoCEv2, MetaRoCE accepts packet loss as inherent to large-scale Ethernet fabrics and compensates via receiver-driven flow control and intelligent NIC processing. This architectural shift enables Meta to build AI clusters on standard merchant silicon switches from Broadcom and Cisco rather than custom interconnect hardware, dramatically reducing cost and deployment time. The protocol's out-of-order packet handling and multipath support decouple it from specific fabric topologies, allowing hyperscalers to optimize for power and space efficiency. Meta has developed a companion conformance suite with Keysight for validation, making this a reproducible standards-based approach that validates the broader industry pivot away from proprietary interconnect toward Ethernet-centric AI infrastructure.

2026 marks the inflection point where the industry transitioned from retrofitting networks onto AI clusters to designing them as first-class infrastructure. While interconnect represents a small slice of total AI infrastructure spend, it remains the dominant source of deployment risk and idle GPU capacity. Scheduled Ethernet and enhanced Ethernet variants (Ultra Ethernet, Broadcom Tomahawk/Jericho, DriveNets' scheduled fabric) are displacing proprietary solutions for scale-out, now reaching 100K+ XPU clusters. The open ecosystem wins on cost and vendor optionality even where it trails proprietary interconnect on raw latency. Ethernet dominated new scale-out AI fabric builds by mid-2025, driven by cost, broad ecosystem support, and multi-vendor availability. When scaling to hundreds of thousands of GPUs, the 10-20% latency penalty of well-engineered Ethernet is vastly outweighed by reduced NIC and switch costs, operational simplicity, and technology optionality.

OpenAI reduced Sol pricing by more than 20% on August 21, 2026, with the promotion running at least through November 21, dropping from $5/$30 to $4/$20 per million input/output tokens. For platform and ML engineering teams evaluating cost-per-outcome: this represents strategic margin compression in frontier models, forcing teams to recalculate inference economics. Sol is part of OpenAI's three-model GPT-5.6 family alongside mid-range Terra and budget Luna, with the price cut signaling that capability parity with Anthropic is now assumed, making pricing the differentiator. Operational impact: teams with long-term commitments should audit their model routing logic—the cost differential between Sol and cheaper alternatives may now justify more aggressive fallback strategies for non-latency-critical workloads.

Ode acquired Casper Studios for enterprise deployment expertise, while former Nvidia researchers raised $90 million for world model startup Veeda AI, and London-based Callosum secured €85.4 million to unify AI models and chips. For practitioners: Ode's acquisition consolidates enterprise deployment patterns under Anthropic's stack, signaling that API access alone is insufficient—enterprises need pre-built skills and integrations. OpenAI's open-sourcing of Harness directly addresses token efficiency, a critical pain point for cost-sensitive agentic workloads. This move puts pressure on proprietary agent frameworks to justify their overhead; if open Harness achieves feature parity for agent orchestration, the differentiation shifts to model quality and governance layers.

Databricks ($5B), Firmus ($2B for Asia-Pacific data centers at $10.5B valuation), Castelion ($1B for hypersonic missile manufacturing), and Etched ($700M) collectively raised $8.7B, representing a shift from software-margin businesses to capital-intensive infrastructure plays. For platform engineers and infrastructure teams: this funding distribution reveals investor conviction that AI's ROI bottleneck is no longer model capability but operational infrastructure—compute allocation, power delivery, and physical colocation matter more than marginal model improvements. Unlike traditional venture capital focused on R&D headcount, these rounds convert dollars directly into capex, explaining why crossover investors like T. Rowe Price and Blackstone Tactical Opportunities are replacing traditional venture funds. Operational implication: enterprises competing on AI will need to model infrastructure costs as commodity-like, with governance and efficiency becoming competitive moats rather than just capability.

Anthropic moved four core agent infrastructure components from beta to production on August 20, 2026. Computer Use shifts from the computer_toolset_20260501 beta to computer_toolset_20260801 GA—eliminating the need for the beta header and enabling multi-action requests in a single turn. The new Browser Use tool extends computer control to web applications specifically. Skills API (/v1/skills) and Files API (/v1/files) both reach GA status, shedding their respective beta headers (skills-2025-10-02 and files-api-2025-04-14). Files API now uses standard GA response format with file expiration times (expires_in_seconds/expires_at) and id-based filtering. For teams building bounded automation: the stack enforces credentials isolation, domain restrictions, and human approval gates at the API level—not post-facto guardrails. Skills API and Files API are also available through Microsoft Foundry, with updated computer and browser use tools coming soon to Google Cloud Vertex AI. The practical implication for SREs and agents running against live systems is that the API surface is now stable enough to pin to GA versions without tracking beta header deprecations every quarter.

Claude Academy exports Anthropic's employee onboarding—the 4D AI Fluency Framework, ever-boarding, and problem-first tutorials—to the public for free, with design emphasis on mindsets and agency over feature checklists, with practice embedded in every tutorial. 26 courses split into four learning paths, organized by level rather than by topic, running from absolute beginner to developer shipping to production. As of August 21, 22 structured courses exist alongside a larger library of 289 total resources; nine of the 22 target AI Fluency across different audiences, the rest cover Claude.ai, Claude Cowork, Claude Code, Agent Skills, MCP and the Claude API. For builders, the interesting move is the Claude Academy Skill—education routed through the same skills architecture used for coding workflows. Access requires only an email; no subscription barrier. For SREs and platform teams, the broader signal is Anthropic positioning AI fluency as infrastructure competency, not just API documentation—relevant for teams standardizing Claude across internal tools.

AT&T deployed LiteLLM model routers that assess task complexity and route queries to appropriately-priced models, cutting costs 56% with only 2% performance decline. This reflects a broader industry shift toward cost-effective AI deployment as telcos scale internal LLM usage. For network operations practitioners, the implication is significant: intelligent model routing allows telcos to preserve performance where it matters most while cutting waste on routine tasks. AT&T is reportedly exploring open-source AI models to reduce Anthropic spending as companies seek more flexible and lower-cost alternatives to proprietary systems. This signals growing pressure on commercial AI providers and suggests network teams should expect telcos to standardize on hybrid open-source and proprietary stacks rather than vendor lock-in.

AT&T invested in hardware startup Hark, collaborating on device certification and network connectivity for AI-native consumer platforms designed for global connectivity. The partnership ties directly into AT&T's strategy to monetize its converged 5G and fiber footprint through AI-driven device and data opportunities, potentially deepening customer relationships if AI-native hardware gains traction on its network. For network engineers, this represents a shift in architecture planning: rather than treating devices as passive endpoints, telcos are now designing network policies, edge compute, and routing around AI inference workloads that require low-latency, deterministic paths. AT&T also completed a $1.10 billion bond offering to fund network infrastructure alongside the Hark partnership, underscoring capital commitment to converged AI-ready infrastructure.

Recent operator-vendor work demonstrates AI-driven optimization in commercial mobile networks: T-Mobile and Ericsson achieved 15% higher downlink throughput and 10% spectral efficiency gains on live 5G Advanced networks versus legacy approaches. August 2026 marks a turning point toward AI-native networks where architecture, from RAN to core, is designed with AI as an intrinsic component; the Open RAN Alliance's July 2026 AI-Native RIC specification codifies embedding AI functions directly into the RAN Intelligent Controller. AI spans anomaly detection, predictive maintenance, alarm correlation, root-cause analysis, RAN optimization, capacity forecasting, energy optimization, customer-experience assurance, security analytics, and network automation, with AI agents increasingly handling multi-step operational investigation and decision support. For practitioners evaluating AI-RAN ROI, the T-Mobile results provide concrete benchmarks for spectral efficiency gains achievable at scale.

Salesforce published the second edition of its Agentic Enterprise Index in August 2026, drawing on aggregated usage data from the Agentforce platform across thousands of businesses plus a proprietary research study of 4,689 respondents from May 2026. Seven in 10 customer-service sessions are now handled autonomously among organizations in its dataset, with escalations remaining steady. Organizations increased activated agents by nearly three times, while the average time required to create an agent fell by 53 percent, with the average agent able to perform six distinct business actions by the end of 2025, compared with two at the beginning of the year. For enterprise practitioners, this demonstrates that agentic AI has moved decisively from pilot phases into operational deployments at scale. Retail, travel, financial services, and the public sector all showed sharp increases in agent activity, though regulated industries tended to deploy more sophisticated systems.