Uploaded September 2026 | Updated September 2026, 8 minutes ago
Nick Kuhn needed to get his lab’s vSAN backend off 10 gig networking. What he
ended up with is a pair of used enterprise Arista switches, two mismatched
firmware versions, and a console cable that did not fit his laptop. An agent
picked the cards, picked the cables, worked out the USB-C converter, flashed one
switch off the other over SFTP and had the config written before the gear
arrived. He has still never logged into either switch.
@thecote and David Zendzian have Nick Kuhn on from Cloud Foundry Weekly, ahead
of Cloud Foundry Summit in Heidelberg, and spend the hour on the operations half
of agentic AI rather than the programming half. Nick explains why Broadcom hosts
every internal MCP server on Cloud Foundry and why running one on your laptop is
the non-approved way, then gets into the unglamorous part: an agent per
foundation, where its memories live when the VM reboots, and what a request
queue looks like when the thing on the other end is a GPU rather than a
database. David pushes on the thing everyone skips, which is whether you would
hand that switch migration to somebody who had never configured a switch. Coté
works out that Nick is running three different modes at once and that only one
of them is vibing.
Key topics:
Why every MCP server at Broadcom runs on the platform, and why “on your laptop” is the non-approved way
What breaks when a load balancer built for a three-second database call sits in front of a GPU doing prompt processing
Where an agent’s memories, credentials and scratch disk actually live once you stop running it on a laptop that goes to sleep
Why you assume every credential an agent touches is already leaked, and the agent that confesses it echoed one
The switch migration nobody logged in for, and the experience that made it safe rather than reckless
Technology is easy and people are hard, except the technology is not easy, it is complex, and that is what an agent eats for breakfast
Model selection as a job: the council that decides what is allowed, and the person who works out what is any good
Who is managing who, when the human is the slow factor
Interested in a platform that can help you with enterprise-y excellence? Check out: vmware.com/products/app-platform/tanzu
Tanzu Catsup is a weekly conversation about platform engineering, cloud-native operations, and building software in large organizations…and, of course, AI.
Check us out Fridays at 10am US Eastern/4pm Amsterdam time! In YouTube: youtube.com/playlist?list=PLAdzTan_eSPSlg3nySSAI7DjrbN2Bt56r
Index
0:00 Nick Kuhn, and what he is taking to Cloud Foundry Summit
3:41 How Broadcom itself uses Cloud Foundry for agentic AI
5:39 Every MCP server on the platform, and the non-approved way
7:41 Running an MCP gateway at Broadcom's scale
10:17 GPUs are not databases, and neither are the timeouts
14:01 Laptop to cluster: MCP for tools, APIs for apps
16:12 An agent per foundation, and where it lives
20:18 The agent's files, the compute, and the storage magic
23:50 Agent memory, swarms and who is allowed to remember what
26:09 Assume the credentials leak
27:26 If you needed it for humans, you need it for the AI
29:06 The 100 gig migration the agent did
32:04 Technology is not easy, it is complex
33:49 Boat ops, and whether you need the experience first
36:28 The platform already knows what you have
38:10 Three modes: vibing, specifying, and managing
40:26 The manager, the purchaser, and the senior engineers
41:28 Talking to agents, and the one ginormous prompt
44:55 Never logged into the switches
45:46 Small local models that feel like the big ones
48:04 Who picks the models: a council, or a platform engineer
50:13 Risk and legal on one side, does it work on the other
52:40 Agents will figure out how to talk to each other
Hosts: @thecote and David Zendzian
Nick Kuhn needed to get his lab’s vSAN backend off 10 gig networking. What he
ended up with is a pair of used enterprise Arista switches, two mismatched
firmware versions, and a console cable that did not fit his laptop. An agent
picked the cards, picked the cables, worked out the USB-C converter, flashed one
switch off the other over SFTP and had the config written before the gear
arrived. He has still never logged into either switch.
@thecote and David Zendzian have Nick Kuhn on from Cloud Foundry Weekly, ahead
of Cloud Foundry Summit in Heidelberg, and spend the hour on the operations half
of agentic AI rather than the programming half. Nick explains why Broadcom hosts
every internal MCP server on Cloud Foundry and why running one on your laptop is
the non-approved way, then gets into the unglamorous part: an agent per
foundation, where its memories live when the VM reboots, and what a request
queue looks like when the thing on the other end is a GPU rather than a
database. David pushes on the thing everyone skips, which is whether you would
hand that switch migration to somebody who had never configured a switch. Coté
works out that Nick is running three different modes at once and that only one
of them is vibing.
Key topics:
Why every MCP server at Broadcom runs on the platform, and why “on your laptop” is the non-approved way
What breaks when a load balancer built for a three-second database call sits in front of a GPU doing prompt processing
Where an agent’s memories, credentials and scratch disk actually live once you stop running it on a laptop that goes to sleep
Why you assume every credential an agent touches is already leaked, and the agent that confesses it echoed one
The switch migration nobody logged in for, and the experience that made it safe rather than reckless
Technology is easy and people are hard, except the technology is not easy, it is complex, and that is what an agent eats for breakfast
Model selection as a job: the council that decides what is allowed, and the person who works out what is any good
Who is managing who, when the human is the slow factor
Interested in a platform that can help you with enterprise-y excellence? Check out: vmware.com/products/app-platform/tanzu
Tanzu Catsup is a weekly conversation about platform engineering, cloud-native operations, and building software in large organizations…and, of course, AI.
Check us out Fridays at 10am US Eastern/4pm Amsterdam time! In YouTube: youtube.com/playlist?list=PLAdzTan_eSPSlg3nySSAI7DjrbN2Bt56r
Index
0:00 Nick Kuhn, and what he is taking to Cloud Foundry Summit
3:41 How Broadcom itself uses Cloud Foundry for agentic AI
5:39 Every MCP server on the platform, and the non-approved way
7:41 Running an MCP gateway at Broadcom's scale
10:17 GPUs are not databases, and neither are the timeouts
14:01 Laptop to cluster: MCP for tools, APIs for apps
16:12 An agent per foundation, and where it lives
20:18 The agent's files, the compute, and the storage magic
23:50 Agent memory, swarms and who is allowed to remember what
26:09 Assume the credentials leak
27:26 If you needed it for humans, you need it for the AI
29:06 The 100 gig migration the agent did
32:04 Technology is not easy, it is complex
33:49 Boat ops, and whether you need the experience first
36:28 The platform already knows what you have
38:10 Three modes: vibing, specifying, and managing
40:26 The manager, the purchaser, and the senior engineers
41:28 Talking to agents, and the one ginormous prompt
44:55 Never logged into the switches
45:46 Small local models that feel like the big ones
48:04 Who picks the models: a council, or a platform engineer
50:13 Risk and legal on one side, does it work on the other
52:40 Agents will figure out how to talk to each other
Hosts: @thecote and David Zendzian
