• Human Infrastructure
  • Posts
  • Human Infrastructure 440: Starlink as Failover, Homegrown Tools, AI Agent Helpline, and More

Human Infrastructure 440: Starlink as Failover, Homegrown Tools, AI Agent Helpline, and More

AuthorAuthorAuthor

Robin Young, Drew Conry-Murray & Ethan Banks
March 19, 2026

THIS WEEK’S MUST-READ BLOGS 🤓

Jack fired up a Starlink Mini as a failover Internet connection at his home, using his Unifi box’s dual WAN connectivity to automate the failover if his primary link goes down. Jack talks through life with Starlink and Unifi, explaining the IPv6 challenges, CGNAT concerns, speeds and latency for him in the UK market. - Ethan

Eva Santos covers device offsets applied during surveys, clever ways to leverage the cone pattern of directional antennas, and the important distinction between co-channel contention and interference. Eva gets right to the point in this post and cross-references industry presentations should you want to dig into these topics more deeply. - Ethan

Gábor Kis-Hegedüs has been writing up a thoughtful storm lately, covering a lot of topics interesting to network automators. This piece presents AI as a kind of computing stack. I think the analogy holds together in interesting ways. He begins…

The more I work with Claude Code, the more I realize we’re not using a chatbot. We’re using something that looks a lot like a computer. Not metaphorically — structurally. The architecture of a modern AI agent maps almost perfectly onto the anatomy of a traditional computer system.

Andrej Karpathy was one of the first to frame it this way: the LLM is the CPU of a new computational stack. The context window behaves like RAM, long-term stores resemble disk, agents act as processes, and tools are invoked as system calls. Once you see it, you can’t unsee it. Every piece of the agent stack has a direct analog in classical computing.

The post expands these ideas from there, calling on other industry presentations and projects to talk through a mental map of AI stack components. A worthy exercise, in my opinion. This helped me reconsider how I’ve been thinking about AI tech in the context of a new framework. I first read this over a week ago, and I’ve been chewing on it since. - Ethan

Tony Mattke shares a list of scripts and other tools he’s written himself to handle small but important tasks. The blog has descriptions of the tools and links to the repository. - Drew 

Tim writes about his experience using Claude Code to write an AI agent.  The agent doesn’t work. Sometimes you can learn more from failure than success, but it can be embarrassing to publicize it, so props to Tim for being willing to share.  

In addition, Tim hopes the community will take a look and offer ideas about where he might have gone wrong. He writes “I’m going back to the drawing board to rebuild, but I’m stuck wondering is this operator error or AI error? Did I miss some obvious workflow item? Some magic words or markdown file that would have made this work? If so how do I know what to do to fix it for next time?” 

If you’ve written an agent or two, or done other coding work with Claude, check out this post and drop Tim a suggestion or two. - Drew 

MORE BLOGS


Modern networks generate massive amounts of telemetry. Most monitoring platforms cope with this scale by being selective or down sampling data over time, sacrificing the detail engineers need for accurate troubleshooting.

Statseeker takes a different approach.

It collects high-frequency SNMP telemetry every 60 seconds and retains it at full resolution, even at large scale environments. This allows network teams to analyse performance exactly as it occurred, detect anomalies quickly, and prove root cause without relying on aggregated data.

With automatic network discovery, intuitive dashboards, and granular historical data, engineers can quickly identify congestion, capacity trends, or abnormal behavior across their networks.

Because Statseeker is self-hosted and built to scale —from smaller environments to some of the world's largest networks, organizations maintain full control of their data across physical, virtual, cloud, and hybrid infrastructure while gaining deep operational insight.

Explore how network teams are using Statseeker to improve performance and reliability:

TECH NEWS 📣 

This is a weird one to pop up. The articles talks about SCION, and it’s not new. The issue isn’t whether SCION is superior or not, or whether continuing to bolt more crap onto BGP to improve the situation is the right answer. To be fair, I’m not clear that SCION would actually be able to replace BGP as a global routing protocol without refreshing my memory on the details of what all it does. But anyway…

The larger issue as we are several decades into networking technologies is dealing with sticky inertia (which would be a great prog rock band name). It’s more or less impossible to displace BGP for global routing. Just like Ethernet was never intended to perform the services we’ve pressed it into, but we can’t really replace it. Just like DNS is being used in ways not originally conceived of, but we do it anyway. Just like IPv4 lives on, probably in perpetuity, despite IPv6 itself showing some gray hairs.

If you’re of the mind that superior technology should just win on its merits (which is my default position even though I’ve been long disabused of the idea), scroll down to the “The foundation nobody wants to renew” section of The Register’s fine piece, which explains why SCION isn’t really a thing except in specific use cases. Standards development organizations, market forces, and the “swapping the engine while the plane is in flight” problem all appear.

Nothing is certain except for death and taxes…and Ethernet and IPv4 and...😅 - Ethan

A detailed look at how Microsoft got its cloud services approved for US government work despite FedRAMP evaluations that found Microsoft’s security controls were insufficient.  - Drew 

I’m tempted to call the deployment of robot dogs “security theater.” Then again, theater can create effective fictions. Executives who dislike employing humans who want things like a livable wage and sick days might be drawn to the idea of data center facilities patrolled by robot dogs, which upload their surveillance videos to an AI that reviews the images. Nary a human employee to be found. And the C-suite lived happily ever after. - Drew  

MORE NEWS

FOR THE LULZ 🤣

RESEARCH & RESOURCES 📒

RootViz aggregates measurements of the root DNS servers taken by globally distributed RIPE Atlas probes and displays them in a Grafana dashboard. There are also dashboards per root server if that’s useful data to you, as well as some other combined dashboards. Overall, a pretty interesting angle on tracking global Internet health. - Ethan

Cal Newport critiques recent articles foretelling the destruction that AI will wreak upon high-skill workforces. He offers a sensible reality check without downplaying the inevitable disruption that’s likely to arise. (Thanks to Matt Oswalt for surfacing this via LinkedIn.) - Drew  

If you’ve ever gone home from work and wished you could’ve spent more time racking servers, cabling switches, and mapping the network, you are a likely candidate for “Data Center,” a forthcoming game that will be available on Steam. The game is launching at the end of March, but you can download a demo now if you didn’t get enough networking in your day job. - Drew   

MORE RESOURCES


70x faster MTTI. 90% reduction in log costs. It almost seems unreal. Kentik shares
practical ways NetOps teams can modernize their workflows with Kentik AI Advisor.
The article highlights examples of how Kentik AI Advisor can accelerate root cause analysis, simplify traffic investigations, uncover patterns in network data, and support smarter capacity planning. By making advanced analytics easier to access, Kentik helps teams move from reactive troubleshooting to proactive operations. If you are exploring how AI can improve visibility, reduce mean time to resolution, and streamline NetOps tasks, this blog walks through seven use cases to get started.

UPCOMING LIVE EVENTS 🍕🍻

A curated list of near-future meatspace events of interest to network engineers. Sometimes a Packet Pusher or two will be there (noted below).

Subscribe to events.packetpushers.net in your calendar software.

MARCH 2026

RSA Conference 2026
23 - 26 March | San Francisco, CA (Drew attending)

AI For Network Leaders Powered By Selector.AI
25 March | New York City, NY (Ethan attending)

APRIL 2026

Wi-Co Finland
9 April | Helsinki, Finland

TORNOG1 | Toronto Network Operators Group
13 April | Toronto, Canada (Ethan attending)

Wi-Fi World Congress USA 2026
13 - 15 April | Mountain View, CA

Wi-Co Philadelphia
16 April | Philadelphia, PA

Wi-Fi Design Day 2026
16 April | London, UK

Wi-Co Toronto
22 April | Toronto, Canada

Wi-Co Frankfurt
29 April | Frankfurt, Germany

MAY 2026

Extreme Connect 2026
4 -7 May | Orlando, FL (Ethan and Drew attending)

NLNAM Meetup 2 | NL Network Automation Meetup
13 May | Alphen aan den Rijn, Netherlands

Wi-Co Brussels
21 May | Brussels, Belgium

(NH)NUG | New Hampshire Networking User Group (USNUA)
27 May | TBD, NH (Ethan co-organizing)

Wi-Co Memphis
29 May | Memphis, TN

Cisco Live US
31 May - 4 June | Las Vegas, NV (Packet Pushers maybe?)

INDUSTRY BLOGS & VENDOR ANNOUNCEMENTS 💬 

Eridu is going after networking for data centers where AI workloads are happening. I just caught this headline in passing. We haven’t been briefed by Eridu (yet). So I’m not clear on what makes an Eridu network special.

But here’s a clue. “Eridu’s network switch delivers the order-of-magnitude advance in performance, radix and efficiency needed to meet current and future AI demands.

With a clean break from the limits and trade-offs of current architectures, Eridu delivers previously impossible capabilities including:

  • Fewer network tiers to reduce latency and network jitter

  • Single-hop scale-up domains with thousands of GPUs

  • Scale-out domains of millions of GPUs”

To me, that implies a fabric that’s not just a leaf-spine interconnect with all of the congestion & oversubscription challenges an architect has to consider in that sort of topology. They’re maybe doing something novel to allow more direct connect GPU-to-GPU communications at massive scale. Perhaps dynamic reconfiguration at L1, like one of the offerings from the folks at Corespan.ai (formerly Drut). Cool stuff you can do with an optical switch and math. But that’s speculation. We’ve reached out for a briefing and will report back if Eridu can make time for us. - Ethan

TL;DR. Cisco announced the Open Transport 3000 Series. “Multi-rail architecture integrates optical components for multiple fiber rails into a single line card, providing exponential improvements in footprint and operational scale for multi-petabit traffic demands.” There are additional notes in the blog about the NCS 1014 system as well as new optics.

The big idea here is about the need to handle more data at massive scale because of AI compute workloads. - Ethan

This is somewhat similar to the Cisco announcement above. “Arista Networks today announced the formation of a multi-source agreement (MSA) for XPO, a revolutionary 12.8 Tbps liquid cooled optics module that supports a front panel density of 204.8 Tbps per open compute rack unit, a 4X improvement compared to 1600G-OSFP optics.”

It’s wild to me just how much data everyone thinks we’re gonna be moving around so that GPU clusters can tell us things. But I guess if we build it, they will come! - Ethan

MORE INDUSTRY NOISES

DYSTOPIA IRL 🐙

TOO MANY LINKS WOULD NEVER BE ENOUGH 🐳

LAST LAUGH 😆