SuzieQ Gets an Agent Brain and Amazon Challenges Human-in-the-Loop
Top 3 Highlights
1. SuzieQ Gains an MCP Interface — Your Network Now Has a Memory and a Voice
Why it matters: SuzieQ already indexed your network state across twenty-plus tables — BGP sessions, OSPF neighbors, routes, ARP entries, MAC tables, VLANs, MLAG, EVPN. The MCP layer makes all of that queryable by any MCP-compatible AI agent without custom code. The emerging pattern is three MCP servers working in concert: SuzieQ for observation, an Ansible MCP server for configuration, and a PagerDuty MCP for incident escalation. Together that's a closed-loop AIOps stack where the observability, remediation, and notification layers are all tool-callable and composable. This is the architecture that turns network troubleshooting from a human detective exercise into a guided, agent-assisted workflow. The pre/post-change validation use case alone — assert network state before change, push config via Ansible MCP, assert again after — is worth the integration effort.
Takeaway: Add the SuzieQ MCP server to your lab environment this week. The path trace tool is immediately useful for connectivity verification, and the time-travel BGP query is the kind of capability that changes how you approach incident post-mortems.
Sourceshttps://blog.ipspace.net/2026/06/worth-reading-suzieq-mcp/, https://github.com/netenglabs/suzieq
2. Amazon VP Says Human-in-the-Loop AI Governance Is Broken by Design
Amazon distinguished engineer and VP of Security Eric Brandwine made a direct challenge to the "human in the loop" consensus at a recent AWS security event, calling it a governance mechanism that fails predictably. His argument: humans are inconsistent, and after repeated false alarms with no consequences, they stop paying attention — a phenomenon well-documented in healthcare, aviation, and emergency services as "normalization of deviance." You design a control, it generates noise, people tune it out, and then the one real event slips through.
Amazon's proposed alternative isn't "remove humans from the loop" — it's "accountability end to end." Every agent deployed in Amazon's infrastructure has an independent identity assigned to it. If an agent writes and executes a script that causes an outage, the engineer who deployed that agent is still the accountable party. The agent's identity traces back to a human, but the human isn't approving every step — they're owning the outcome.
Why it matters: This is a direct counter-narrative to the "always require human approval" posture that many security teams have defaulted to as they grapple with agentic AI. Brandwine is not arguing for no oversight — he's arguing that point-in-time human approval is a worse governance model than agent identity traceability plus accountability ownership. For network automation specifically, this maps directly to the question of whether every Ansible push needs a change-ticket approval or whether a well-scoped agent with a traceable identity can operate within defined change windows autonomously. The answer is starting to look like the latter, but only if your identity model can answer "which agent did this and who owned it."
Takeaway: Audit your current AI/agent governance model. If the control is "a human clicks approve," ask whether your team still reads those approvals carefully after six months of false negatives. Agent identity with accountability ownership may be the more durable architecture.
Sourceshttps://www.theregister.com/security/2026/06/20/why-amazon-hates-human-in-the-loop-ai-governance/5258639, https://thenextweb.com/news/amazon-human-in-the-loop-ai-governance-normalization-deviance
3. Microsoft Locks In 2.67 GW of Natural Gas for Twenty Years — The Power Lock-In Era Is Here
Microsoft signed a twenty-year power purchase agreement with Chevron for 2.67 gigawatts of dedicated natural gas generation in West Texas. The project, called Project Kilby, is the largest power deal between a US oil major and a Big Tech company on record. Chevron will invest $7 billion to build a new gas plant using GE Vernova turbines; first power is targeted for 2028. The deal is designed as dedicated, behind-the-fence power — Microsoft's datacenter takes the output directly, minimizing grid impact.
Why it matters: This is a structural bet, not an operational one. Twenty years is longer than most network gear lasts. The hyperscalers have concluded that grid interconnection timelines (eighteen to twenty-four months for utility approval, longer for major capacity) cannot keep pace with AI infrastructure demand. The answer is to own the generation. What's worth watching is the precedent: Chevron, an oil major, is now in the AI datacenter power business at a scale that dwarfs most utility projects. The site selection calculus for any large deployment is shifting — available generation capacity is now a first-class variable alongside fiber, land, and latency. The FERC cost-causation shift covered last week means grid interconnection is getting more expensive for everyone else at the same moment the hyperscalers are bypassing it entirely.
Takeaway: If you're evaluating colocation or datacenter sites at any significant scale, add "path to dedicated generation" to the due diligence checklist alongside interconnect timelines and utility rate structures.
Sourceshttps://www.theregister.com/on-prem/2026/06/22/texas-lassoes-massive-microsoft-datacenter-and-20-years-of-gas-turbine-emissions/5259764, https://www.chevron.com/newsroom/2026/q2/chevron-signs-20-year-power-agreement-with-microsoft-for-west-texas-data-center
Network Automation
NVIDIA Publishes Telco Agentic AI Autonomy Architecture
NVIDIA's developer blog published a detailed autonomy model for telco networks this week that has implications well beyond telecom. The framing uses TM Forum's autonomous network levels taxonomy — most operators sit at levels two and three (predefined solutions, selective domains). Reaching level four and five requires agents that can understand operator intent, sense the network in real time, research and develop plans, weigh tradeoffs, and coordinate governed actions across domains. The post argues that the constraint is no longer model quality — it's whether operators have built an autonomy platform where agents can draw on a shared stack of domain models, policy controls, tools, and digital twins.
That platform framing is the signal. A single agent calling a single API is not level four. Level four requires the agent to have access to real network state (digital twin), know what it's allowed to do (policy controls), understand the domain it's operating in (domain models), and be able to coordinate with other agents handling adjacent functions. All four simultaneously. Most "agentic NetOps" announcements in 2026 are delivering one or two of those, not all four.
Takeaway: Before evaluating agentic network operations tools, map which of the four platform pillars they actually deliver versus which they assume you've already built. Missing any one of the four is a production blocker.
Sourceshttps://developer.nvidia.com/blog/how-telcos-build-autonomous-networks-with-agentic-ai/
Extreme Networks Launches Agent One for Enterprise Network Management
Extreme Networks announced Agent One this month, framing it as a new class of agentic AI for enterprise network management rather than a chatbot layer on top of existing dashboards. The platform includes Extreme Exchange, a skills marketplace that functions as the control plane for the agent. Agent One Coworker (the first tier) ships in July and handles guided workflows. Agent One Operator, due in Q4 2026, is designed to run scheduled workflows and respond to events autonomously without continuous human oversight.
The architecture — a skills marketplace that agents pull capabilities from — mirrors what we've seen from Itential with FlowAI and the broader MCP server pattern. The skills abstraction decouples the agent's reasoning capability from its domain knowledge, which means you can update what the agent knows how to do without retraining the underlying model.
Takeaway: Add Extreme Networks Agent One to the evaluation list alongside Itential FlowAI and Datadog Bits AI if you're scoping autonomous network operations. The differentiation will be in the skills breadth and the policy guardrails, not the underlying model.
AI / ML
MCP Recursive Composition: Servers Connecting to Servers
The June 2026 MCP specification update is expected to formalize "server-as-agent" capabilities — MCP servers that can themselves connect to other MCP servers, enabling recursive composition patterns. This has been implemented informally in production deployments for months, but the spec formalization matters because it standardizes authentication, capability discovery, and loop detection. The Obot platform has been demonstrating enterprise MCP gateway patterns including recursive composition at recent developer summits.
Why it matters: Recursive MCP composition is what turns a collection of point tools into an actual agent network. SuzieQ MCP + Ansible MCP + PagerDuty MCP isn't three separate integrations — it's a composable observability-remediation-notification pipeline where an orchestrating agent can call each as a tool. The formalization in the spec means vendor implementations will converge on the same patterns rather than each building bespoke multi-hop logic.
Sourceshttps://www.itential.com/resource/guide/the-ultimate-mcp-guide-for-network-automation/
Datacenter & Infrastructure
New Datacenter Developments: Kansas, Virginia, Africa, Europe
Digital Realty announced a 600MW campus in De Soto, Kansas, expanding US hyperscale footprint into the Midwest alongside Kansas City's existing power and fiber advantages. In Virginia, developer Takanock filed for an 87-acre, 180MW campus in Strasburg near I-81 — designed with closed-loop cooling and dedicated power generation, avoiding grid interconnection dependency. In West Africa, Kasi Cloud opened Lagos's first carrier-neutral AI-capable hyperscale-ready campus, and EdgeConneX plans three Italian campuses totaling approximately three billion euros.
The pattern across all four announcements: dedicated or behind-the-fence power, closed-loop cooling, and sites chosen for fiber accessibility over proximity to existing power grids. Strasburg in particular — off the well-worn Northern Virginia corridor — signals that secondary Virginia markets are now viable once you bring your own power.
Takeaway: "Bring your own generation" is becoming a site selection requirement, not a nice-to-have. Model it into your capital planning from the first feasibility pass.
Sourceshttps://www.datacenterdynamics.com/en/news/digital-realty-plans-600mw-campus-in-kansas-acquires-investment-firm-columbia-capital/, https://www.datacenterdynamics.com/en/news/87-acre-project-tallmadge-to-be-built-in-strasburg-virginia/
Science & Emerging Tech
Atom Computing and Nu Quantum Target Distributed Quantum Networking
Atom Computing and Nu Quantum signed a memorandum of understanding on June 17th to develop distributed quantum computing architectures by linking multiple neutral-atom quantum processing units via photonic network switches. The focus areas are qubit-photon entanglement interfaces, photonic network switches, and distributed fault-tolerant computing protocols. The partnership combines Atom Computing's neutral-atom QPU expertise with Nu Quantum's photonic interconnect technology.
The significance: the quantum computing community is increasingly converging on a distributed architecture thesis — connecting multiple smaller QPUs via optical links rather than scaling a single monolithic processor. This mirrors how classical HPC moved from single large systems to clusters connected by high-speed interconnects. The quantum networking problem is genuinely hard (maintaining entanglement fidelity across photonic links at scale), but the architectural direction is now consistent across multiple groups.
As a callback to last week: Duke and IonQ's three-node entanglement demonstration (covered Monday) and the Atom-NuQuantum interconnect partnership are different parts of the same convergence. One proves the physics works, the other starts engineering the infrastructure.
Takeaway: Track quantum networking entanglement rate improvements as the leading indicator for when distributed QPU architectures become practical. The current ~0.1 entanglements per second needs to reach roughly 1 per second before distributed computation becomes competitive with monolithic scaling.
Sourceshttps://quantumcomputingreport.com/atom-computing-and-nu-quantum-form-transatlantic-alliance-to-network-neutral-atom-qpus/, https://www.prnewswire.com/news-releases/atom-computing-and-nu-quantum-partner-to-unlock-utility-scale-quantum-computing-302801484.html
The Fun One: Six Weeks Hunting a Ghost in Rust
Cloudflare spent six weeks chasing a race condition in the hyper HTTP library — the open-source Rust library that underpins their Images service running on every machine in their edge network. The bug: transformation requests returned HTTP 200 with no errors logged, but the image data was silently truncated. A response that should have been two megabytes arrived with a few hundred kilobytes. No error. No log entry. Just shorter data.
The race condition only triggered under specific conditions: concurrent requests above a certain concurrency threshold, for images above a certain size, after the architecture change that created a more direct local connection between the Workers runtime and the Images service. Six weeks. One race condition. In an open-source library that thousands of projects depend on.
The detail worth sitting with: the bug was discovered via Cloudflare's own testing, fixed upstream in hyper, and they published the full root-cause writeup. That's the open-source trust model working as intended — a major edge network operator finds a subtle bug, fixes it publicly, and documents it for everyone who shares the dependency.
Takeaway: If you're running any Rust HTTP services with hyper, check your version and review the Cloudflare writeup. The concurrency threshold conditions that triggered this race are not unique to their architecture.
Sourceshttps://blog.cloudflare.com/hyper-bug/
Quick Takes
- Project Glasswing on Hold? Or Not? Packet Pushers NB580 covers conflicting signals on whether Anthropic's Project Glasswing critical infrastructure vulnerability program is pausing or continuing amid the Mythos export control situation. The export control saga continues to have downstream effects. (Sources: https://packetpushers.net/podcasts/network-break/nb580-project-glasswing-on-hold-or-not-why-you-should-hold-in-person-background-checks/)
- Hyperscalers Sign White House Ratepayer Protection Pledge: Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI signed an agreement to fund grid upgrades and not socialize datacenter interconnection costs to general ratepayers. An attempt to preempt further FERC intervention after last week's show-cause orders.
- HPE Spaceborne Computer 4: HPE and Kioxia are sending an AI inference compute module to the Moon later this year. The storage subsystem revealed by ServeTheHome is more interesting than the marketing suggests. (Sources: https://www.servethehome.com/this-is-the-storage-kioxia-hpe-spaceborne-computer-4-bringing-ai-compute-to-the-moon/)
Watch Today
- Extreme Networks Agent One Coworker shipping in July — first chance to evaluate the skills marketplace architecture in a vendor product.
- FERC grid operator responses due in August — PJM and MISO cost-causation plans will set market expectations for AI datacenter interconnection pricing for the next several years.
- MCP spec June 2026 update — recursive server-as-agent capabilities moving toward formalization. Watch the official MCP GitHub for the release.
- Atom Computing / Nu Quantum collaboration — the entanglement rate milestone to watch is 1 per second. Current demonstrated rate from Duke/IonQ last week was 0.095.
Morning Briefing — June 23, 2026 | Amaze Networks | beestonlabs.dev
Get the briefing in your inbox.
One email per weekday morning. Same writing, same sources — no audio required.