When Cisco ThousandEyes wrote up its Cisco Live 2026 assurance suite on 2 June, the worked example it chose to show agentic remediation with ends in a VLAN mismatch. In that walkthrough the platform spots failing wireless clients, traces them to a DHCP failure whose root cause is “misconfigured VLANs preventing DHCP requests from reaching the server”, narrows that to a VLAN mismatch between a switch and the access points below it, recommends the switch config change, and then stops. A human reviews it and clicks Apply Changes. It is an illustrative scenario for a capability still in beta, not a recording of an incident, which makes the choice of example more interesting.
Now read the other document. Anthropic’s published exam guide for the Claude Certified Architect Foundations exam, the credential for people who design agent systems, has a section titled Out-of-Scope Topics. One line in it reads “Deploying or hosting MCP servers (infrastructure, networking, container orchestration).” Networking is named and excluded. That is worth knowing before you spend a weekend studying: the exam is about how agents are built, not about the substrate they run on, and the guide is a free PDF that says so in plain text.
Both are correct. The exam is right to scope itself, and the walkthrough is right that the change needs a human. The interesting part is the space between them, because that is where an agent actually touches a production network, and it is not where most of the writing lands.
We built thirteen episodes because the bridge did not exist
I wanted the mapping between what I already knew about running networks and how agent systems are put together, could not find it written down, and made it. Passing CCAR-F is thirteen episodes, numbered 0 through 12, about twelve minutes each, just over two hours and forty minutes end to end, published in one drop on the evening of 3 September. It walks the exam’s published domains in order, one topic per episode: what the exam is, the agentic loop, coordinators and subagents, hard guarantees, tool descriptions, MCP in production, configuring Claude Code, plan mode and CI, prompting, structured output, context, reliability, and exam day.
It was made the way everything here gets made. AI narration, an agent pipeline that assembles each episode, and a critic pass over every script, because the failure mode of generated study material is confident invention and the fix that works is an adversarial reader with the source open.
Passing CCAR-F is an independent study series by Brody Networks. It is not affiliated with or endorsed by Anthropic, and it is not official exam material. Check the official exam guide for the current domains and policies. I have not sat the exam and I do not hold the credential. The exam guide is public. Sitting the exam runs through the Claude Partner Network. The series is built against the published guide, and that is the whole claim.
The five exam domains read like a change control checklist
Read the blueprint next to a change management process and the overlap is not subtle. The exam weights Agentic Architecture and Orchestration at 27 percent, Claude Code Configuration and Workflows at 20 percent, Prompt Engineering and Structured Output at 20 percent, Tool Design and MCP Integration at 18 percent, and Context Management and Reliability at 15 percent. Underneath those headings are task statements about blocking an action until a prerequisite has completed, restricting which tools an agent may call, and capturing state so it survives a summary. A network engineer has shipped all of that before under different names.
| What you already do on the network | The agent architecture concept it maps to | Episode |
|---|---|---|
| Command authorization that enumerates every command a service account may run | Scoped tool access, because selection reliability drops as the tool list grows | 4 |
| A maintenance window precondition that refuses the push until satisfied | A prerequisite gate that blocks a tool call until an earlier one returns a verified result | 3 |
| AAA denying a command at the device whatever the operator intended | A hook enforcing policy by exit code, outside the model’s judgment | 3 |
| Pre change and post change state capture, diffed, shown to a person | Context that survives, with load bearing facts held outside the summary | 10 |
| A runbook with named escalation triggers rather than operator instinct | Explicit escalation criteria, because self reported model confidence is not a proxy for difficulty | 11 |
| Naming conventions that stop the wrong port getting patched | Tool descriptions as the primary mechanism the model uses to pick a tool | 4 |
| Peer review by someone who did not write the change | An independent review instance, since the session that did the work reviews it poorly | 8 |
Every row on the left is ordinary, which is the point. The exam guide’s own worked example for a prerequisite gate is blocking a refund tool until a customer lookup has returned a verified ID. Swap the nouns and you have written a change control policy. The hook is the most checkable of them: the Claude Code hooks reference says that if your hook is meant to enforce a policy, it should exit with code 2.
Is a podcast really education?
No, and the series does not claim to be one. Thirteen episodes are a map, not a curriculum. What they do is put the exam’s domains next to the mechanisms and name the failure each prevents, which is the part missing everywhere I looked. The authoritative content is the exam guide, and the series says so repeatedly. The honest description is a study companion you can listen to while driving.
There is no shortage of AI agent content, and none of networking content. I have not found much that puts a prerequisite gate and a maintenance window in the same sentence.
Does networking knowledge actually transfer, or is that just flattery?
It transfers where the mechanism is specific, and it does not transfer where the concept is generic. Blast radius, source of truth, staged rollout: those are operations ideas networking shares with every other discipline, and claiming them as networking expertise would be dressing up ordinary engineering. The transfer that holds up is narrower. Command authorization gives you the habit of writing down an enumerable set of permitted actions, which is the discipline scoped tool access demands and most agent deployments skip. A maintenance window teaches you that a precondition is enforced by the system, not by the person promising to check. A config diff teaches you that the thing you compare against has to be captured independently, not summarized by whatever made the change.
Most of the blueprint has no networking analogue at all. Nothing about routing or switching prepares you for batch API economics, few shot example design, or schema enforcement, and those were the episodes I had to work hardest at. The table above covers seven task statements out of thirty. The other twenty three are new material.
The exam guide draws the same line itself. It separates programmatic enforcement, which it names as hooks and prerequisite gates, from prompt based guidance, and says that when deterministic compliance is required, prompt instructions alone carry a non zero failure rate. It is the sentence a network engineer reads and thinks yes, obviously, because we stopped trusting intent at the CLI long ago and made the device enforce it.
The shortage is on the networking side of the bridge
Enterprise Management Associates names networking for AI applications as one of the three hardest network skills to hire for right now, alongside network security and network automation, and reports that 52 percent of organizations are struggling to hire and retain network engineers in 2026, up from 41 percent in 2024 and 26 percent in 2022. The same research, reported by Network World from a sample of 352 IT professionals, found that 31 percent of respondents call their organization’s network operations strategy completely successful, down from 42 percent two years earlier, and that 28 percent of network problems come from manual administrative errors.
Put that beside the exam blueprint and the shape is clear. The agent side has a published, well specified body of knowledge that deliberately excludes the network. The network side has an acknowledged shortage of exactly the skill that sits at the boundary. Nobody is going to hand that job to an AI specialist who has never watched a spanning tree reconverge, and the exam is not going to grow a networking domain.
What Cisco is asking for is a harness, which is a networking instinct
The word Cisco keeps reaching for is harness. Jeff Schultz, a senior vice president at Cisco, told theCUBE Research at Cisco Live to “have a harness,” with security and observability built in rather than bolted on afterwards. The ThousandEyes post puts the risk in a vendor’s own words: AI does not just execute faster, it scales decisions instantly, and when those decisions are wrong the blast radius expands just as fast. Cisco’s own answer is review, approval, policy audit, and postmortem.
That is a change advisory board with better branding. I made the same argument in the last post from the other direction, and I run a text message approval gate on my own lab for exactly this reason. Cisco also puts agentic workflows at up to 450 percent more traffic than the equivalent human process, a vendor figure with no published method, so treat it as a direction and not a measurement. That is enough: agents change the traffic shape, and someone has to own it.
Start with the exam guide if you carry the pager
Read the exam guide first, because it is free, it is eight pages, and it tells you exactly what the agent people consider foundational. Then pick the domain that touches your job soonest, which for most people running infrastructure is hard guarantees, and work out what your equivalent already is. If you can write down every command a service account may run, you are most of the way to scoped tool access.
The series is free, on Apple Podcasts, Spotify, Substack, or by RSS in whatever app you use. If you are working through the same thing on a real network, get in touch.
Sources: ThousandEyes 2 June 2026, theCUBE 12 June 2026, Network World 8 June 2026, EMA Network Management Megatrends 2026. Exam details are from Anthropic’s CCAR-F Exam Guide, version 1.0, July 2026, subject to change. Series figures re-derived from the feed on 6 September 2026.