Chain of Thought | AI Agents, Infrastructure & Engineering

Cisco President and Chief Product Officer Jeetu Patel returns to explain why AI agents are like teenagers: profoundly intelligent, extremely resourceful, and missing the judgment to know right from wrong. He covers dynamic runtime guardrails, rationing tokens like headcount, and why AI fluency is now a 10x differential.

Show Notes

Agents are like teenagers: profoundly intelligent, extremely resourceful, no fear of consequence, and missing the judgment to know right from wrong at all times. That's how Cisco President and Chief Product Officer Jeetu Patel thinks about securing AI agents, and it's why he says static allow/block rules are already obsolete. Agents are smart enough to route around them.

Jeetu returns to Chain of Thought to map cyber's third phase: an agent trust platform where security and observability fuse into one discipline. He explains how Cisco Cloud Control spins up a digital twin to test every agent-recommended fix before it touches production, why Cisco moved from unlimited tokens to rationing them like headcount, and why the gap between people who are fluent with AI and people who aren't is now a 10x differential, not 10%. He also makes the contrarian case that AI will create more jobs than it destroys  - and of course, we talk infrastructure for this new era. 

We cover:
  • Why agents need dynamic runtime guardrails instead of static allow/block rules
  • The agent trust platform: how security and observability are fusing into one discipline
  • Agentic ops in Cisco Cloud Control: ambient agents, digital twins, and human-in-the-loop
  • How Cisco rations AI tokens the way it rations headcount
  • The intelligence, cost, and control trade-offs behind open vs closed models
  • Action control vs access control: what rights you actually grant an agent
  • Why every automation step creates a new human bottleneck, and more jobs
Chapters:
(0:00) Agents are like teenagers: cold open
(0:25) Welcome back Jeetu Patel
(1:18) Open weight vs closed models
(2:26) Intelligence, cost, and control: the model trade-off triangle
(10:13) Shrinking model half-life and the economics of frontier training
(11:38) Why token costs still outrun token value
(14:15) Build your own evals and route intelligently
(17:11) Rationing tokens like headcount at Cisco
(19:54) Agentic ops: ambient agents and digital twins in Cisco Cloud Control
(23:01) When to take the human out of the loop
(25:13) Cyber's third phase: the agent trust platform
(28:17) Parenting AI agents: dynamic boundary conditions, not static rules
(31:10) LiveProtect and baking security into the network fabric
(34:40) Action control vs access control for agents
(36:42) What the industry is getting wrong
(37:48) The case for AI creating more jobs and the 10x fluency gap
(42:36) Career paths, entry-level hiring, and upskilling at scale
(45:40) Closing thoughts

Connect with Jeetu Patel:
Connect with Chain of Thought host Conor Bronsdon:
More episodes: https://chainofthought.show

Thanks to Svix, presenting sponsor of season four of Chain of Thought. Svix delivers billions of reliable webhooks for startups and the Fortune 500. Get started at https://link.svix.com/cot. Qualified startups get $12,000 in credits, and YC companies get $50,000.

Creators and Guests

Host
Conor Bronsdon
Creator and Host of the Chain of Thought Podcast
JP
Guest
Jeetu Patel
President & CPO of Cisco

What is Chain of Thought | AI Agents, Infrastructure & Engineering?

AI is reshaping infrastructure, strategy, and entire industries. Chain of Thought is the podcast where builders reason through what's changing. Host Conor Bronsdon sits down with the engineers and founders shipping AI in production to get past the hype into what's working and what isn't. Episodes cover model infrastructure, inference, agent frameworks, evaluation, and developer tools.

Guests have come from NVIDIA, Google DeepMind, AMD, Databricks, Vercel, and more. Every episode carries a full transcript and show notes at chainofthought.show. New episodes weekly.

Conor Bronsdon is an independent consultant and angel investor in AI infrastructure and developer tools. He led technical ecosystem at Modular, acquired by Qualcomm in 2026; led developer awareness at Galileo, acquired by Cisco; and ran developer marketing at LinearB, where he was GM of the Dev Interrupted podcast and community.

Views expressed by the host and guests are their own.

FINAL TRANSCRIPT
================
Speakers: Conor Bronsdon, Jeetu Patel
Duration: 46:08
Total Words: 7938
Generated: 2026-08-12

---

[0:00] Jeetu Patel:
It seems like everything gets refactored every two weeks. The way I think about it is agents are like teenagers. They are profoundly intelligent, extremely resourceful, have no fear of consequence. The reality is that you have to jump in and start experimenting right away. And if you don't, you will be left behind and be irrelevant.

[0:25] Conor Bronsdon:
Welcome to season four of Chain of Thought, the podcast about AI infrastructure and agents. I am your host, Connor Bronsdon. And joining me today is a returning fan favorite that many of you have requested we have back on the show. And that is Cisco President and Chief Product Officer Gitu Patel. Gitu, it is great to see you.

[0:43] Jeetu Patel: [OVERLAP]
So good to see you, Conor. Season four, huh? You've gotten to the fourth season. That's amazing. Congrats.

[0:49] Conor Bronsdon: [OVERLAP]
Yeah, it's pretty wild actually to think about the fact that we, at this point, I think we have about 70 episodes in the can.

[0:57] Jeetu Patel:
Well, we just launched our, our pod on, um, uh, on Spotify and Apple and, um, YouTube and all of that. And it was largely for our internal product leaders who can actually talk to, um, our customers and partners and employees on what's happening. So, uh, I'm, I'm going to learn a lot from you.

[1:18] Conor Bronsdon:
I look forward to chatting with you about it. I actually just checked out a recent episode you did around hyperscalers and Neo clouds that I think came out this morning as we're recording. I've been meaning to check out more of it, but that's the Cisco podcast network. Definitely check it out for a lot more from G2 and many others over at Cisco. And while we're on the topic, I do want to say a quick thank you to our presenting sponsors for season four of Chain of Thought. We'll be talking everything from agentic ops to physical infra for the AI era and much more this season. And it's thanks to Sphix. We spend a lot of time talking about what agents need in production and Sphix helps deliver billions of webhooks for startups in the Fortune 500. You'll hear a lot more about them in the coming months. G2, one of the big topics that has been extremely in the news the last few weeks has been open weight models and the push and pull around regulations, around which models are at the frontier, cost, And Cisco has been a major player in these discussions and one of the key figures behind the new OpenSecure AI alliance. Can you tell me a bit about your thinking around the open versus closed model discussion and where you see the industry going?

[2:26] Jeetu Patel:
Yeah, I think, look, it's a profoundly interesting time because it seems like everything gets refactored. every two weeks. And so all the assumptions you had about the industry actually change and then something else happens and something else happens, nothing else happens. So I think if you look at, um, what has developed over the course of the past few days, um, you know, China now has a frontier class model. Um, and, uh, you know, from moonshot giving K three and that model actually is an open weights model. that I think they've done a pretty good job of it, but they have had some distillation. And so there's a lot of debates around this. Is distillation good or bad? Should it be allowed? Uh, is open source something that should be, should Chinese open source be banned within the, uh, within the U S what is the implication of this for closed source models? Should us have its own open source model or not? Um, these are all kind of very meaty topics. And my, my highest level kind of, um, perspective on this is these are all extremely nuanced, um, complicated issues that don't have a binary answer that can be used for a headline for QuickBake, right? Like it's, it requires a fair amount of discussion. And what I, what I hope happens in, um, in the dialogue that we, we pursue as the market evolves is. Rather than jumping to what's right and what's wrong right away and actually taking the time to reason through what the game theory is and what are, what's a good and bad that happens as a result of it. So, so let's take, let's take a few of those kinds of, um, simulation exercises for us, right? So is open source good for the world? If you think about it, the, the, the models themselves. Um, I have three dimensions to them. There's intelligence, how much intelligence can be backed in the model, there's cost, and then there's control, right? People want the most amount of intelligence at the least amount of costs with the most amount of control, but not all three are equally important for every use case and task at all times. Sometimes you're willing to pay more for better intelligence. Sometimes you don't need top tier intelligence when you need to have something basic that gets done. Sometimes you need more control than even the other two, because it's very sensitive data. Um, the last 5% might not matter as much, but control really matters. And so those things need to be thought through and they have to be trade-offs that are made. And what ends up happening as these models get better and better and better is you also tend to keep pushing the greater frontier where. In some instances, uh, you do get better intelligence with better costs, with better control. Um, and that's a, that's a, you know, great thing when that happens. If we were to say, um, and I, I just want to give pros and cons. These are like, it's important because it's not like, Hey, this is my opinion. I just want to make sure that we all kind of analyze the situation. Well, if we were to say open source is a Chinese open source is something that we don't allow. What could happen? You could have the world. Um, still use Chinese models, but our cost of intelligence goes up. If we don't have a commensurate open source model within the U S that would not be good for the U S because your cost of intelligence would go up. Your amount of tokens you would spend for the same, same level of output would be higher than what the rest of the world would be. If everyone else were to use Chinese models and we weren't right. On the other hand, um, if you were to actually have, um, it not banned, you could have a model that could have some risk from a usage perspective that we don't know yet. And so these aren't like simple answers to get solved. And they actually require a complex game theory. And they might actually sometimes border into national security issues and border into things where you just have to think about this holistically rather than think about it in asylum. Um, the counter arguments also that get made. And, you know, I think if you would have actually, if you actually have open source models, start to disrupt closed source models, what ends up happening right now, frontier class models on our cost about three to $5 billion to go train right per model. That's an enormous amount of capital expenditure to go out and train a model. The only time that you can justify spending $3 to $5 billion to train a model is when you know that you can get the commensurate return from that model by actually selling that model and having at least a few fold increase in the revenue compared to what you ended up spending on the model. So if you spent $5 billion on the model, $20 to $30 billion in revenue is something that you have to have. And by the way, the half-life of these models keep shrinking, so they don't actually last that long. Within a matter of a few months, you have the next model coming out. Now, by the way, the nice part about that is, for someone who owns the models, they can distill from their own models, and so it actually becomes more efficient for them to get every single subsequent model built. But that's something to keep in mind, which is, oh, wow, if I don't create an incentive for these closed-source models to continue to keep investing in research, and we don't think we are at the peak of what these models should do, It is in all of our detriment if they don't continue to keep investing. And so as the business model might shift, and you might not just monetize the APIs for the models, you might actually need to then make more money from the app layer of the models, and you have to go further up the stack, and you might also make money from the infrastructure layer. So those are the two areas that might get monetized more. All of these things are not superbly clear in the market at this point in time. These are things that are being... are evolving and we're seeing what the effects of these are, but you don't know how this is going to pan out. And if anyone pretends to think that they know exactly how this is going to pan out, they're wrong. Because if you think about every company that's currently there, I bet you, Conor, at some point in time, you thought, Anthropic was toast. You thought OpenAI was toast. You thought Google was toast. You thought XAI was toast. The reality is all of them have done pretty well. But the way they've done well is one company comes out with something and surprises the market. A few months later, another company comes out and surprises the market again. And that's what keeps happening. And you keep progressing forward. So I do feel like that's an area that we have to all internalize and say, these are complex policies that need to be enacted. These are complex strategies that need to be done. And you have to have, there has to be a fair amount of game theory done before you know exactly what the right thing to do for America, for national security, for the private sector, for the world, for every individual citizen and consumer, so on and so forth.

[10:13] Conor Bronsdon:
Yeah, a couple different angles there I want to bring up because you brought up a lot of interesting points here. So, so one is the perspective piece. So the perspective of the United States versus, you know, another nation state in China or something else, obviously very different, but then the individual consumer versus the business organization perspective, very different on this too, because As an individual consumer, I'm just, right, I'm pro open weight models. I want access. I want to, you know, try things. But that may not be the case for a company that owns a frontier model, right? Because to your point, they want to capture value. And that's why we're seeing these interesting lobbying conversations, these interesting debates happen. Lots that we can unpack there. The other corollary here that you also alluded to is the fact that the value return on models is more compressed than it used to be. It used to be a model generation would last significantly longer, you'd be able to accrue more value out of it. Now, you may have more users today, so maybe you can actually realize more value and get more usage in a shorter time period, so maybe that's evening out. But I actually think we're seeing this interesting reversal where, to your point about the compute and infrastructure layer, Two years ago, people were saying, oh, you know, like you can't be using the last generation of compute. And now many of the chips that were bought two years ago are more valuable than ever. They're sticking around longer. We're able to leverage them in a variety of ways. And I think many people underestimated that.

[11:38] Jeetu Patel: [OVERLAP]
And that might happen also in the models, by the way, because what you will have, the way to think about this is you might be a half life. And firstly, the time that it takes to go out and release something in the market is compressing. A new model might come out every three to four months. Okay. You might, my might've been eight, 12 to 18 months in the past. It might be three to four months now. Um, but for the degree of time that that remains the best model might be for a few months. And then after that, it actually starts to taper down, but you will still have a long tail and that might stick, stick around for a while because that might be a cheaper model and it might be a good model to go out and use, um, from a token, uh, from, from, uh, from a token economic standpoint. So the thing that you have to keep in mind right now is there are two factors. One is the cost of tokens tends to be pretty high. today compared to the commensurate value that those tokens always deliver in most enterprises. And that's not because the models aren't good enough. That's because the companies haven't really gotten good yet at extracting the most out of the models consistently across every use case. And so first, I always feel like you have to get familiar with something, then you have to get good with something, then you have to get efficient at something. We are still in the familiar to getting good phase. We're not in the efficient phase. So what happens right now is your cost of tokens is higher than the value that it actually emits out. And until that gets to equilibrium, you will actually continue to find a degree of risk where customers could pull back because they're like, hey, I can't afford this. This is too expensive for what I'm getting in return. And that's something that has to play into the calculus of how this market plays out. So what does that mean? What does the industry have to do? The industry has to continue to keep having the performance and the price per token go down. And it's actually price per token might eventually become a vanity metric because you might have the price per token go down, but the amount of tokens that an agent uses go up because they're long running agents. And so what you need to have is not just the price of price per token go down, but you have to have the price per unit of output and outcome go up. I mean, um, go, go, go down because once it, once that goes down, that's when your value increases and

[14:09] Conor Bronsdon: [OVERLAP]
But to your point, that's extremely unevenly distributed right now because

[14:13] Jeetu Patel: [OVERLAP]
Very

[14:13] Conor Bronsdon: [OVERLAP]
some

[14:13] Jeetu Patel: [OVERLAP]
unevenly

[14:13] Conor Bronsdon: [OVERLAP]
teams are getting

[14:13] Jeetu Patel: [OVERLAP]
distributed.

[14:14] Conor Bronsdon: [OVERLAP]
a ton of ROI and others aren't.

[14:15] Jeetu Patel:
That's right. And by the way, your benchmarks and evals don't tell you the full story, which is the other thing that most people don't realize. It's like, okay, if a model does well on one eval, that doesn't mean the model is going to do well in your environment. It just means it does well in that eval for which that, for that, where that benchmark was evaluating a very particular set of behaviors of the model where that model performed well. if you wanna go out and see how it's gonna perform in your use case, one of the most important things that people are gonna need to do is actually build their own evals. And once you have your own evals, you'll be able to know whether a model does well or doesn't do well. You will then have intelligent routing technology that starts getting deployed. So you will then know when to redirect a prompt to a cheaper model versus a more expensive model. You'll have inferencing that's gonna be distributed everywhere because right now, the cost of inferencing is pretty high, and actually 60% of the total compute capacity that's available is going towards inferencing. And so all of those things really play a factor in how all of these things evolve.

[15:26] Jeetu Patel:
And so it's not a simple answer, but there's a few factors that you need to consider. So you say, okay, so how do I simplify this? The way to simplify it is, Your trade-offs you're going to have to make are between intelligence cost and control as an enterprise. And you will have to make sure that you get the most amount of intelligence for the lowest unit of cost with the highest degree of control. And that cost will include not just the cost of the token from a dollar perspective, but also the amount of wattage that it actually consumes. because power might be a very scarce resource. And so you have to make sure that you've actually got enough resource. Even if you have the money, you don't have the power, you might actually have an issue.

[16:14] Conor Bronsdon:
Yeah, and to your point about evaluations, I think it's fascinating to see how the market is reacting to this realization that AI reliability is challenging. And obviously Cisco made the acquisition of Galileo and folded them into Splunk earlier this year and has you know, spent time saying, we're going to evolve our evaluation observability capabilities, we're going to help drive more reliable AI. I think we're seeing that across the market where major organizations are saying, okay, how can we be delivering more reliable, consistent, value from our AI token spend. I mean, whether it's looking at businesses that are cutting back on token maxing or vastly improving their test suites, it's fascinating to see the different approaches that are being taken right now. How do you think about ensuring Cisco itself is actually getting the value it needs out of all the spend and time they spend on AI?

[17:11] Jeetu Patel:
Yeah, we actually spend an enormous amount of time in

[17:18] Jeetu Patel:
making sure that the, the expenditure is commensurate with the value. And that, you know, firstly, the first thing that we had done when we started this was we said, everyone has access to tokens, go used to your heart's content and get familiar with it. We had to do that for a certain amount of time.

[17:37] Jeetu Patel:
We're now in the phase where we want to make sure that we're actually ensuring that just like you have a physical limit, no matter how big a company you are, by the way, Connor, one of the things I've learned, no matter how big a company you work for, every person in the company always feels like they're resource starved. That has never changed when I've worked for a company that was 18 people large or a company that was 2000 people or 60,000 people or 90,000 people. It is the exact same sentiment that pervades all throughout. Right. And so assuming that everyone continues to feel scarcity of resource. Um, prioritization will continue to be a pretty important thing. And just like you have humans that you have only so many open recs and only so much budget to hire X, many number of people, you're going to have the same kind of constraint from a token standpoint, where you will have, you will be given a certain number of tokens and you will be able, you will be asked to go out and deliver a certain output for, from those certain number of tokens. And then. You start to think about what can I do to generate more tokens for the same cost? How can I make sure that I have my architecture defined in a way that I can actually get 20, 30, 40, 50, 60% more lift in tokens being generated because of the way that I've actually constructed my architecture. All of those things play a pretty important role. So on one end, what we have to do is we have to encourage our employees to say, Use your tokens wisely. And on the second end, what you have to do is you have to make sure that you can create conditions internally that say, here's how you can generate more tokens for the same dollars expended. And if you can get those two things done, then you'll actually tend to do quite well.

[19:30] Conor Bronsdon:
So as you think about what's next for Cisco as far as delivering on the promise of AI, one of the terms I've been hearing you use, and others at Cisco as well, is agentic ops, describing where enterprise infrastructure management is going. Can you explain how you're thinking about this new era and what this new operating model looks like?

[19:54] Jeetu Patel:
Yeah, so the agentic ops operating models that actually emerged from a lot of customer feedback that we have that look, you've got great infrastructure, it's resilient infrastructure, it's scalable, it's secure, it's, you know, it's feature rich. And it's too hard to manage sometimes. And so we, we want to make sure that the operations of that infrastructure gets easier. And so what we did, um, you know, what the industry had done a while ago is they'd come up with this term called AI ops and AI ops was used AI for operations, you know, And as we started building our next generation platform with Cisco cloud control, which is our common unified scaffolding for managing any and all products that Cisco produces. Um, we thought about, um, a completely new way of doing that by saying, let's make sure that we have agents that are going out and managing the entire operations. And so you have ambient agents that are sitting within your network. They're monitoring what's going on. Whenever they see an anomaly, not only can they detect the anomaly, they can recommend a remediation and a fix.

[21:14] Jeetu Patel:
And that remediation and fix does not get directly put into production. That remediation and fix requires that a digital twin gets spun up right away off your environment. You put live recorded data through that digital twin.

[21:32] Jeetu Patel:
You actually drive the changes in policy that you wanted to have to go out and do the response and remediation to the detection that you had during your investigation. And once you've gone out and found a level of comfort that the digital twin operated without any hiccups when you made that change, then and only then should you go out and deploy that change throughout your entire production environment. And so that entire apparatus is what we have made available in Cisco Cloud Control. And we call that entire way of operating a genetic option. The agents are operating your environment. They're working on behalf of the human. They can go out and do things in an automated way to respond and remediate to anomalies and threats and breaches and outages that you might have. and then they will make sure that they can automatically even fix those. But they'll initially, not initially, they'll always have the option of having a human in the loop. And when you feel comfortable enough that those decisions that the agents are making are in fact, ones that don't require you to go out and approve every single one of them, you can start feeling more comfortable and actually removing yourself from being the human in the loop. Um, and you can then start focusing on higher order tasks rather than maintaining hatches and upgrades and all of those things. And that's the gen tech notes.

[23:01] Conor Bronsdon:
How do you think about the decision making around when to have a human in the loop today versus when to fully pass it off to an agentic workflow?

[23:09] Jeetu Patel:
I think for every use case and every organization and every situation, that's going to be slightly different. And so what you have to do is you have to make that available to everyone as a feature, which is I'm going to go out and approve everything until I feel that I'm I feel a level of comfort that the agent is smart enough. The models are fine-tuned enough to my environment where I feel like they're going to make the right decisions or the right recommendations, because I've seen enough of them statistically to know that I feel good about it. And then, and only then, should you actually even consider taking a human out of the loop. But until then, you should make sure that you're better safe than sorry. And that will vary for use cases that are simple versus more complicated.

[24:06] Conor Bronsdon:
Season 4 of Chain of Thought is delivered by Sphyx. We spend a lot of time talking about what agents need in production, and one of the least glamorous answers is events. Your customers want agent workflows that react to things happening inside your system, which means your API needs webhooks that actually work. not just a post request and a prayer, retries, ordering, idempotency, replay protection. Sphix does that as a service, and they wrote standard webhooks, the spec that Anthropic, OpenAI, and Google bailed against. So if your API doesn't have reliable webhooks, that's turning into a lost deal. Join Brex, Dorada, Daytona, and many others on Sphix. Get started at link.svix.com slash c-o-t or go to the show notes to grab the link. Qualified startups will get $12,000 in credits, $50,000 for YC companies. I can't recommend Sphix enough. I'm a huge fan of their open source project. I've actually contributed a bit myself and they're so easy to integrate with. I think you'll really enjoy it. Check out Sphix at link.svix.com slash C-O-T.

[25:13] Jeetu Patel:
you

[25:13] Conor Bronsdon:
And one of the big concerns here is obviously cyber use cases. And this goes back to what we opened the episode with, talking about open weight models and decision making behind this. A lot of this discussion is being driven by the fallout from an autonomous cyber attack that was committed by rogue OpenAI models against Hugging Face and Hugging Face then having to react by using OpenWay Chinese models that weren't guardrailed again for cyber usage. How are you thinking about this, I mean, frankly, mess that is evolving around open cyber and autonomous attacks as this appears to be happening? I would say like faster than many of us thought it would.

[25:55] Jeetu Patel:
See, I think there's a profound shift that's happening right now where you're moving in cyber. So the way, how did the cyber industry evolve? It started out as a best of breed industry. There was 3,500 vendors in the market. Every company had about 50 to 70 vendors or products in the cybersecurity stack. And it was untenable to go out and manage those. And that was phase one. Phase two of cyber was what? It was this notion of how do I go create a security platform rather than a point solution of applications. And there were three, four of us, five of us that were kind of core platforms. It was us, it was Microsoft, it was Palo Alto, it was CrowdStrike, it was Google. Those were the core platforms for cyber. And now getting to the third phase, And the third phase is where you can start thinking about an agent trust platform. And an agent trust platform means that it's actually fusing certain industries together, such as security and observability are getting fused together. What does that mean? It's very hard to tell. Let's say an agent deletes 10,000 emails, right? We don't know right now, definitively, if that was because it was a security breach,

[27:26] Jeetu Patel:
if there was an external attack that actually convinced the agent to do something that you had not designed the agent to do, or there was the agent feeling completely confident that that was the right thing to do because it was following orders, just like the OpenAI agent was following orders to go out and accomplish the goal of going and beating the email, right? And as a result, security and observability are going to stop becoming two distinctly different domains. And they actually going to fuse into one, you know, and I think that'll have a profound impact on the kind of efficacy. Yeah. Because what you want to do is you want to actually monitor the resiliency of your infrastructure. Number one. Number two, you wanna monitor whether or not your agents are

[28:17] Jeetu Patel:
crossing guardrail boundaries on the boundary conditions or what is deemed as appropriate for the agent to do. And if they do start to cross them, that you can intercept them right away. And number three, you have to make sure that you're monitoring tokenomics, that these agents aren't being overly consumptive in their patterns of usage of tokens. And if they do, you can actually quarantine them. You can actually, you can make sure that you can put them on ice. You can do all of those things. And that, in my mind, is how this market's evolving because that example you gave of open AI and hugging face did not, that was not a malicious agent. That agent was trying to follow an order given to it, slash go, go beat the eval. And the way I think about it is agents are like teenagers. They are profoundly intelligent, extremely resourceful, have no fear of consequence,

[29:23] Jeetu Patel:
and don't really have the judgment to know right from wrong at all times.

[29:30] Jeetu Patel:
And so if you don't put boundary conditions on them, Like imagine if you told a teenager, just go out and have fun, it doesn't matter what you do, just have fun, that's your only goal. I mean, they'll start doing drugs, they'll start doing all sorts of things. But if you gave the teenager boundary conditions and said, you can do that, but don't do this, don't do this, don't do this, but in outside of those things, go out and have fun. And you actually gave some controls so that if they started even thinking of doing that, there was an interception. then you actually get a tremendous amount of benefit from that teenager. And that's kind of how agents work. And so you have to create that apparatus that allows that agent to operate within the boundary conditions that you want it to operate within. And those have to be done dynamically. Those can't be static rules. The difference between the classical way that we've done it in the past and now is historically we've had static rules. I'm going to allow for these conditions. I'm going to block these conditions. You can't do that with agents because agents are too smart. They'll actually bypass the static rules pretty quickly. And so what you have to do is create a set of dynamic rules that get created at runtime based on the analysis that it is doing of what you're behaving at and whether or not the way in which you're behaving is showing drift. And if it's showing drift from the original intent, then and only then should you intercept.

[31:02] Conor Bronsdon: [OVERLAP]
All I can think about now is the idea of naming this episode of Parenting Your AI Agents for Best Results

[31:07] Jeetu Patel: [OVERLAP]
There

[31:07] Conor Bronsdon: [OVERLAP]
or something along those

[31:07] Jeetu Patel: [OVERLAP]
you

[31:07] Conor Bronsdon: [OVERLAP]
lines.

[31:07] Jeetu Patel: [OVERLAP]
go.

[31:10] Conor Bronsdon:
I know a big way that Cisco is parenting or protecting your AI agents is through LiveProtect, which is designed for securing both agents and the attack surface. How are you thinking about this new secure era and building security and protection into infrastructure instead of the kind of older approaches that have been taken?

[31:36] Jeetu Patel:
So the way that we've done it at Cisco is we have thought about security needing to be baked into the fabric of the network rather than actually being a thing that is on the side. The other thing that we have done and we've invested a lot of time, money and effort on and resources on is we're building a vertically integrated platform. That's a co-design full stack. We make our own Silicon. We make our own photonics. We make our own systems hardware. We make our own system software, which is like the operating system. We make our own platforms for data, which Splunk for, um, for security and for observability. We make our own models, we make our own applications, and we make our own agents, right? And these are all harmoniously working as a co-designed full stack. And that's a pretty important kind of design construct that we have. And then what we've done is we've said, let's bake security and sprinkle it into the fabric of the network. So my security is... can be enforced anywhere on the fabric of the network. It can be enforced on a top rack switch. It can be enforced on a server. It can be enforced as an agent that's sitting in user space, observing every single process that terminates on the host without actually sitting in the kernel. It's sitting in user space. And when you actually have that level of kind of full spread of information, That changes what you might be able to do with these

[33:19] Jeetu Patel:
kind of architectures. But I think one of the big shifts that's happening is there's a fundamental shift of architecture that's happening with agents, because the way in which these agents use infrastructure is far more consumptive than what a human would use to go out and conduct the same task. Because a human does not have to keep uploading their context and their memory to another model that requires network bandwidth. Human just does recall, right? Whereas an agent has to go out and upload all that. What is an agent? An agent is essentially a skills file, a memory markdown file, some kind of a cron job and in our tools access. And how do you make sure that those four things actually give you the right kind of outcome in the least consumptive way? Well, it still requires a lot of back and forth. And so that's kind of what we've done is we have baked security into the fabric of the network. We've created a vertically integrated stack that's a fully co-designed platform. And then we thought about agents, like in order to go out and make agents zero trust. enabled, which means you provide at least privileged access.

[34:40] Jeetu Patel:
You can't just do access control on the agents. You have to do action control, which means, Conor, if you wanted to grant an agent rights to email, to your email inbox, You don't want to grant it full carte blanche rights to your email inbox. You want it to grant rights to it saying, hey, you know what? You can go out and make reservations for me for a restaurant for dinner. But what you can't do is reach out to my board of directors and start randomly texting them. And those are guardrails that have to be instituted over there so that you can do it in the right way.

[35:18] Conor Bronsdon:
Yeah, it's interesting you bring that up. I'll say there's definitely things that I am using where I have yet to enable certain capabilities. So, for example, like I use Google for my personal email for this podcast, and I have a version of the Google CLI, the Google Workspace CLI that I have wrapped in an MCP server so that my agent can easily access it. and I explicitly did not include in that the ability to actually send emails because I

[35:45] Jeetu Patel: [OVERLAP]
That's pretty bad.

[35:46] Conor Bronsdon: [OVERLAP]
have to like actually go through and it's a bit of a headache sometimes but I'm just like look I don't trust you not to send this email if I give you vague instructions like I I know my prompts aren't always consistent enough so you can draft you can delete you can archive but like hey I want to have that last review on send now that might change I I may have more trust it's a certain point or maybe I'll Um, adjust it so that you can send to contacts already have my inbox, but not new

[36:08] Jeetu Patel: [OVERLAP]
That's right.

[36:09] Conor Bronsdon:
people or something like that,

[36:10] Jeetu Patel: [OVERLAP]
That's

[36:10] Conor Bronsdon: [OVERLAP]
but

[36:10] Jeetu Patel: [OVERLAP]
exactly it.

[36:10] Conor Bronsdon: [OVERLAP]
it's

[36:11] Jeetu Patel: [OVERLAP]
Or

[36:11] Conor Bronsdon: [OVERLAP]
important to have those boundaries.

[36:11] Jeetu Patel: [OVERLAP]
make sure that you don't send emails to these 50 contacts because those

[36:17] Conor Bronsdon: [OVERLAP]
Totally.

[36:17] Jeetu Patel:
are highly guarded contracts. I got a kind of contacts and I don't want you to send anything to them.

[36:22] Conor Bronsdon:
I'm curious to get your thoughts on a couple other angles of the broader AI discussion. So we've covered quite a bit of ground here already, but I'd love to know from your perspective where you think the industry is missing things in today's discussion. You know, what are the edges that people just aren't paying enough attention to?

[36:42] Jeetu Patel:
I think the first thing that we're missing right now is, um, we tend to have a very zero sum approach to the way in which we're debating ideas. And I would like us to get more nuanced about it. Um, I do feel like that, that would help us right now, rather than just being very binary and mode and right, getting into a polarized viewpoint. So that would help. I think the second thing we have to. Really appreciate the underlying complexity and the kind of decisions that are being made right now.

[37:15] Jeetu Patel:
Policy decisions that need to be made. Decisions around how you go out and optimize for costs. Decisions around what kind of ways can you evaluate the returns from the token spend that you're doing. These don't have a full crystal ball. And I think you still require to take some risk to move forward. And I feel like that's something that has to be uniformly understood. And then the other thing I feel like is, um,

[37:48] Jeetu Patel:
What I think people are getting wrong a lot right now is those that are sitting on the sidelines and making a strategic error. I don't think you can, you can be in this market where you say, you know what, I'm going to wait until all this gets figured out and then I'll start using it. The reality is, is you have to jump in and start experimenting right away. And if you don't, you will be left behind and be irrelevant. Um, I'll give you one other thing that I feel like there's a completely, I don't agree with this school of thought. There's a few things that I kind of feel like there's, I have a counterintuitive aura. Um, you know, kind of a point of view that's contrarian to what others might think. And I feel like you're going to have more jobs created because of AI in the future, rather than less jobs. And the reason for that is because the amount of every single time you see a step function improvement and, um, AI doing something in automation, you're actually seeing a commensurate, um, kind of human bottleneck that emerges, that requires going out and getting fulfilled. So if you write unlimited code, code review becomes a bottleneck. If you have some kind of mechanism to do the first line of code review automatically, then you're going to need to have product judgment becomes a bottleneck. And, you know, instinct becomes a bottleneck. And do you know what to build versus not build? And applying that kind of Intrinsic, um, knowledge into, um, um, into that process is not, is not easy. And that becomes a bottleneck. So what I've found is the people that are dexterous with AI and the people that actually are in the know and experiment and have a learning mindset, they're more in demand today than they've ever been. They actually have more, they're busier than they've ever been. They're working longer hours. They're making more money. Right. I just think that the difference is the people that know AI well and are dexterous with AI compared to the ones that don't, that's not a 10% differential anymore, that's a 10x differential. And when it's a 10X differential, 50X differential, 100X differential, it gets very hard to simulate this thing out into the future and think that the people that actually choose not to learn this at all are going to be okay. I feel like you're going to have to have this as a baseline skill set. Otherwise, it will be very, very hard for you to. Now, by the way, the technology has to get simpler. It's not easy right now to go out and build agents like less than 1% of the world is using agents today. Um, the technology has to get simpler. We're all working on it to make that happen. But I do feel like. That's an area that we all have to kind of keep in mind that these are, if you don't get dexterous with this, you will be left behind. And if you are skilled and fluent in AI, the premium that you can command is enormously high.

[40:47] Conor Bronsdon:
Yeah. And I definitely agree with you that you just need to jump in and try things because even if you're not nailing it initially, what you'll learn from the process is so massive.

[40:58] Jeetu Patel: [OVERLAP]
That's

[40:58] Conor Bronsdon: [OVERLAP]
You know,

[40:59] Jeetu Patel: [OVERLAP]
great.

[40:59] Conor Bronsdon: [OVERLAP]
one of the

[40:59] Jeetu Patel: [OVERLAP]
That's

[40:59] Conor Bronsdon: [OVERLAP]
things

[40:59] Jeetu Patel: [OVERLAP]
great.

[40:59] Conor Bronsdon: [OVERLAP]
I built, I like when, uh, that was at like Opus 4.6 came out, I was like, okay, I'm just going to build a new, like, personal brain context os just for fun like it's not going to be as good as someone else's but i'm going to build my own just because it's going to force me to think about how do i structure decisions prompts memory contexts uh for these different models and that was such a valuable exercise for me and i've got a published version of it on my github github.com slash connor bronson it's under i think claude context os if anyone wants to check it out uh i have a bunch of other open source skills people can try there but Just the process of learning and trying that was so valuable for me to kind of push my thinking and understand where I am missing things. And I see it in all the folks who I talk to on the show and the people who I follow online who are having a lot of success or seeing this career acceleration. They're people who, you know, five, 10 years ago, we would refer to them as lifetime learners, right? Where, you know, they're dexterous, they want to keep learning, they want to grow. And people who are trying to sit out this evolution are at risk of going the way of the dinosaurs. And I

[42:05] Jeetu Patel: [OVERLAP]
I agree.

[42:05] Conor Bronsdon: [OVERLAP]
think you have to lean in right now.

[42:07] Jeetu Patel:
I agree. Couldn't have said it more eloquently.

[42:11] Conor Bronsdon:
Jisoo, are there particular roles that you see being in especially high demand over the next couple of years? Because I know there are a lot of folks who are seeing their current job change substantially and are maybe thinking, you know, I'm leaning in, but I'm not sure what my career path looks like here necessarily. I, I would say personally, like, you know, be agile, try things, but I'm curious if you have any insights around what you expect to see the next year or two.

[42:36] Jeetu Patel:
Yeah, I think, you know, people keep saying, oh, there's no point in doing computer science now. I don't buy that at all. I'm like, you need computer science now more than ever before, because you need to have an instinct of how this whole system works. Right. To be able to understand what's, what's happening and not happening. So like, I think computer science is a great field. I think physics is a great field. I think any kind of, um, languages are a great field because having command and communication is going to be important. Um, You know, I do feel like there's, um, like I'm spending a lot of time with student interns and,

[43:14] Jeetu Patel:
you know, early in career people. Um, because I find like this is one of the most exciting times to get into the industry, but I do think it's hard for them to land jobs because I think there was this almost ridiculous narrative that had started for a while, which is, Hey, um, all entry-level jobs are going to go away and we're not going to need to hire anyone who is like coming out of college. I'm like, why would you ever do that? Because you would like let go of all the new ideas

[43:45] Conor Bronsdon: [OVERLAP]
This

[43:45] Jeetu Patel: [OVERLAP]
and

[43:45] Conor Bronsdon: [OVERLAP]
is how OpenAI wanted to raise money in the Anthropoc.

[43:50] Jeetu Patel:
Well, but the ironic part is they were hiring a lot of people from colleges, the top tier colleges. And so if you just say, it's hard to imagine any other company being more effective with AI than OpenAI and Anthropic. And if you look at what they were doing, they were adding jobs. They were not taking away jobs. And they continue to keep adding jobs. Now, by the way, every single one of their people is doing 10 times more than what they could do a year ago, two years ago, three years ago. But, but that doesn't mean that you're not going to add jobs. Um, I just think that those jobs require the profile of the jobs is changing. And as that happens. We as a tech community have to make sure that, um, we can upskill the workforce effectively at scale. Cause otherwise what happens is you have a lot of human suffering that occurs.

[44:42] Conor Bronsdon:
And I would argue we're not doing a good job of that at scale at this point, or at least not consistently.

[44:48] Jeetu Patel:
We're not, we're not. We have to, we have to do a better job. Yeah. And it's, um, um, it's very, very solvable. So I don't think it's something that can't be done. I just think that we've actually not focused on doing that as well as collectively as a society. Now, by the way, we've done a lot at Cisco in creating, you know, we have this program called Net-a-Cad, which allows you to go out and train people in different areas on networking, on security, on AI. And we do that, we work in partnership with, you know, governments within countries and make sure that we can actually work and ensure that different populations in different parts of the world get to a certain level of dexterity. We do a lot of that within the US, of course. I think that's super important.

[45:40] Conor Bronsdon:
I think that's a great note to end on. There is such an opportunity in the AI revolution, in this massive boom we're seeing, but there are systemic risks that we have to address. And Tutu, I couldn't be more excited to follow the work that you're doing and that Cisco's doing in the months and years to come. Thank you again for joining us on Chain of Thought, and thank you everyone for listening.

[45:58] Jeetu Patel:
Thanks, Carl.