It’s been 6 months since the last episode, and it feels like everything has changed. Bret and Nirmal catch you up on the increasing pace of AI change, how it will likely affect DevOps engineers and platform builders, and what’s coming on the podcast. Check out the video podcast version here: https://youtu.be/NkqAMAIW5ks 😇 My new GitHub Security workshop has launched! A free 2-hour workshop with hands-on labs to harden your repos and your workflows from common supply chain attacks. I'll cover how attackers are getting in, and then we'll lock down a sample repo so you know what needs to be done to protect your code. You'll leave with a deep understanding of risks and mitigations as well as a list of helpful tools to keep your repos safe, including my new "gasa" tool for scanning your repos and orgs. ★Show Links★ Agent Harness Skills https://agentskills.io Video of Gradel CEO "You can write code faster, can you deliver it faster?" https://www.youtube.com/watch?v=VOXNJ9_sHuM Is your team still hand-chiseling code? https://www.geocod.io/code-and-coordinates/2026-01-21-hand-chiseling-code/ Creators & Guests Bret Fisher - Host Beth Fisher - Producer Cristi Cotovan - Editor Nirmal Mehta - Host You can also support this podcast by subscribing to my YouTube channel and my weekly newsletter at bret.news! Grab the best coupons for my Docker and Kubernetes courses on Udemy. Join my cloud native DevOps community on Discord. Grab some merch at Bret's Loot Box Homepage bretfisher.com (00:00) - ADOP - Feb 26, 2026 (02:23) - Introduction (03:00) - Season Two Kickoff (04:34) - Productivity Debate and Two AI Camps (06:48) - Ops View from Dev Retreats (10:05) - Automation Debt and App Tsunami (18:21) - Frontier Models Changed Everything (24:21) - Red Pill vs Blue Pill Skepticism (27:23) - Three Focus Areas for DevOps (29:43) - GitHub Copilot and Agentic Workflows (33:09) - Harnesses That Run Longer (33:43) - Skills As SOPs (36:22) - Guardrails For App Tsunami (38:22) - Agent Orchestration a nd Kubernetes (39:42) - Shadow IT Goes AI (40:22) - Inner Loop Diagram Breakdown (42:46) - AI Code Review Councils (45:01) - Context Layer And Docs Debt (51:15) - Career Waves to Agent Ops (59:21) - Future Guests and Episodes (01:02:37) - ★ Support this podcast on Patreon ★
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Nirmal Mehta: non-technical users starting to use these tools to build business applications that aren't devs. Right? They don't know anything about CI/CD. They don't know anything about infrastructure. They don't know anything about code and they're gonna start building apps. if you thought Shadow IT was a big problem back when cloud was new, oh boy. Shadow IT with gen AI is like, that's a whole brave new world that we're entering into.
Bret AI July 2025: Welcome to the Agentic DevOps podcast, and I am one of your hosts, Bret Fisher. My co-host Nirmal Mehta is back in the studio. His studio, not my studio, and we are recording what we're calling season two. It's actually been almost six months since our last episode, and a lot has happened and we quite frankly, have just been overwhelmed. I think a lot of us are feeling this right now. The last three months since December have been a total shift. I'm looking at this as more of the epoch of AI is good enough now, that's
the summary of this entire podcast. Where are we last left off in August, we were talking about AI as a helpful assistant, but yes, you have to spend a lot of time with prompting and managing hallucinations and all that. I'm not really doing much of that anymore, and for the last three months, I haven't experienced much of any, uh, what I would call significant hallucinations. I certainly have seen it make wrong decisions and do things I didn't intend. But it's, it's amazing how much better these are. Truly Opus, and some of these newer models like GPT 5.4 and 4.5 and 4.6 of Opus and
Sonnet have fundamentally changed the game for me and a whole lot of us online. So we talk about that. We go through some of the headlines of the last six months, but we're now starting season two, and we've already got multiple other episodes lined up around the world of AI and its intersection with engineers doing automation and DevOps like things. So, I'm excited. Uh, we've got a whole bunch of shows. I mean, we've never had seasons before, but why not Season two. Let's, let's call it that. And, I'm looking forward to getting these episodes out to you. So this is the first of many and we go pretty random and quick
through a lot of the content we're gonna talk about in this episode. So here we are, 2026. Everything AI. Bret: Hi, Nirmal. Welcome Nirmal Mehta: Hi, thank you. I'm Nirmal Meha, and I'm a principal specialist solution architect and containers tech lead at AWS and, longtime friend and co-host, with Bret on all the different shows that Bret has created over the years. I'm super excited to be back here for the next quote unquote, season
of the Agentic DevOps podcast. Just a quick note. The opinions that I will be sharing on the show are my opinions and not of my employer, Amazon Web Services. And with that. Boy, do we have a long list of things to catch up on? Bret: we can't even catch up on six months. Ok, Nirmal Mehta: No. Bret: so if this is your first Agentic DevOps podcast, welcome to the show. Nirmal Mehta: Yeah. Welcome. Bret: And Nirmal and I started this podcast a year ago, next month. We were in London, I believe, London Cobe Con, CloudNativecon.
Nirmal Mehta: Correct. Bret: We started it there. It felt like the right time to start it. Nirmal Mehta: It really did. Bret: We were talking about it, but it felt like murmurs, like everyone in the industry was focused on how to run AI. Nirmal Mehta: Yeah. Bret: Using AI, and agents were new. The agents were a new idea. And MCP was less than six months old at that point. Nirmal Mehta: Yeah. Bret: I think we talked about that first one. Like we mentioned A2A, like a brand new protocol for agent. Agent or Nirmal Mehta: Correct. Yeah. century ago. Bret: I know it feels like forever.
Well, none of that matters anymore. We're gonna get into it. But we had some shows over the summer. We had some guests, we had Laura Tacho on. We Nirmal Mehta: Yeah. Bret: Solomon Hikes on talking about sort of the world of CI in DevOps, in AI. Like how that relates to AI and how they see it. They have since pivoted. We might talk about that. But this is gonna be the kickoff episode. So I wouldn't call it a reboot. Nirmal Mehta: Yeah. Bret: I wouldn't call it Nirmal Mehta: a Bret: comback. Nirmal Mehta: No, Bret: I've been here for years. that's the LL Cool J quote. just for the kids that don't know that 80's, 90's music. and so let's back up a second and remember where we were.
Our last episode was Laura Tacho in really around August. I think we aired it in September. We recorded maybe around August Nirmal Mehta: About, Bret: of AI not being a a productivity booster for business when it comes to business goals, business outcomes of the software teams, and how AI wasn't actually trailblazing this amazing improvement in business productivity. It helps with coding and speeds up coding, but that's a small portion of a developer's week and that is only one part of the software lifecycle.
So we kind of had some sort of bad news around, hey, maybe five or 10%, you know, business productivity. And then third quarter happened, and third quarter we got more news. we saw more studies saying, yeah, AI has had, I think the Yale study, which is debatable. I don't think everybody agreed with it, but I think the headline there was AI has had zero effects on jobs so far. So there are definitely different camps. You've got one camp on the red pill side that is describing it as 10x developer
and just amazing productivity benefits. But everyone that I know that says that is an independent developer. Nirmal and I were at a great conference together and this conference was a group of friends that are all builders, they're all SaaS makers, developer consultants, making products, working for big and small companies sometimes, but they're all highly invested in using AI for their job. Nirmal Mehta: Yeah. Bret: One thing that Nirmal and I come from is we come from the ops world
where these people tend to come more from the pure dev, pure builder world. Nirmal Mehta: Yeah. Bret: From enterprise teams and big consulting companies and complication and legacy and monolith, like that's our world Nirmal Mehta: Yeah. The operations side. Yes. Bret: I do have my friends and I do spend that time making, I am an independent creator. I am a solo developer and I have solo developer friends. And lemme tell you, these two camps are completely different when it comes to AI.
Like it's two sides of the same coin on AI, but they're completely different experiences and opinions based on the size of the team and the size of the company. What do you think Nirmal Mehta: And, and there's like, there's a lot there. for me, it was not only attending, that solo developer retreat with you, Bret, which was amazing. but I followed that up, by attending the Pragmatic Summit in San Francisco, just a few days later. Bret: back to Nirmal Mehta: of this recording, that was just two weeks ago. And, from my perspective, you're right Bret, both of us are kind of,
our experience and our expertise is grounded in operations system and administration, SRE DevOps, engineering, DevOps practices more on the Ops. Right? I would say like we're lowercase dev, big case Ops people, Bret: That's a Nirmal Mehta: and we've. Bret: Yeah. Nirmal Mehta: We've spent our careers bridging, you know, we're DevOps practitioners, right? First and foremost. And we've been bridging this tension that's been between the
developers and the operations folks. And of course, those lines have gotten very blurry with cloud native tooling and, the modern organizational structures these days. in my day to day, I typically talk to platform engineers, operations folks, engineering teams that are, like SRE teams. I don't talk to developers, especially not solo developers much. And so for, for me, for me, Bret: contracts and enterprise Amazon contracts.
Nirmal Mehta: for me, these two events back to back were kind of like an expedition from me into this. The world of developers, like the current state of developers, it was like a exploration of that world that I, I'm not usually in. And it was fascinating and I kind of agree, Bret, like we've been working for over a decade on bridging DevOps together culturally, right? Containerization, cloud native automation, and making developers lives easier and
also making operations folks' lives easier and bridging that tension, right? that classic throat over the fence. I think you're kind of right with respect to Agentic systems, AI. kind of reigniting some of that tension because the adoption, what I saw, and granted this was definitely a group of leading edge adopters of this technology for sure. But what I saw there was, as a developer, the adoption and the usage
of these tools, at least in some part of the SLDC lifecycle, right? The inner loop development was pretty phenomenal. I just saw really amazing things, and just productivity boosts. I don't want to guess on the percentage, but very clear, direct re, you know, outcomes there. again, this is the Age Agentic DevOps podcast and our mission on this podcast, one, is we wanted to focus on the. More of the operations side.
How are these AI and agentic systems affecting the operations? Like everything kind of from when the code is built, to like the testing CI/CD and then actually running a production and then the operations of those applications in production. And it was interesting to see how the acceleration of these tools, these tools accelerating application development was now highlighting in very stark terms the need for organizations to have automation from the CI/CD all the
way through to their infrastructure in the same way that we've been talking about for the last 20 years. Because now there's a wave, a tsunami, I think of hundred x more applications gonna be built. starting now. The wave has already started. And I don't think as DevOps practitioners, as operations folks, I don't think we're ready to handle this. wave of new applications and this new way of software development that's happening.
I think there's like even more tension, but the fundamentals are still the same. That's the takeaway I got is like you now more than ever, you organization needs automation. It needs solid DevOps practices, it needs solid documentation and you know, efficient operations and I'll say automation again and likely we'll need to adopt agentic tools. We'll talk about this a little bit more, but I think. We will have to adopt agentic tooling to even just operate these systems at the
scale to meet the needs of these, the tsunami of applications that are coming. Bret: Yeah, I mean, it really talks, you know, it talks, this leans into like, is your CI/CD debt? We talk about code debt, Nirmal Mehta: Yeah. Bret: but ops, debt, SRE debt, DevOps debt, these are all same, similar problems. And if your systems are not enabled enough that a developer can spin up a new idea, iterate on it, put it into staging, put it into production without
your involvement, if they're not enabled. We might label that as platform engineering, but if they're not enabled for that, you're suddenly gonna have a lot more work on your plate if you're having to deal with everything that the developer's creating. ' cause if the developer can code 10 x faster, let's just say they can, or three x whatever, two x, it's gonna be faster than today. They're coding more. We know that the negatives around like the business productivity, that all starts to fall down because the code is only a small portion of the problem. Like we, they create code now. We dev people, DevOps people have a new problem, which is they're shipping more
updates to the same code and they're probably gonna be shipping a lot more new repos as they spin up things. And this is what we saw at our, retreat before you went to San Francisco. Is that a solo creator can now be working on three or four projects at the same time, or a solo dev in a team can now, you know, things that before, well that would be a couple of weeks of work and I'd have to get approvals. Assuming that you don't have to go through approvals and all these different processes, you can iterate on an idea very quickly, especially
in those first early days of an idea. And what I'm finding as a DevOps engineer is I'm creating way more tooling for myself. And so let's go back to the third quarter because we're gonna step through a couple of major shifts in Nirmal Mehta: Yeah. Bret: paying attention that have enabled all this. Nirmal Mehta: Yes. Bret: enable it for the DevOps engineers, not just the developers. Nirmal Mehta: if you remember, some of our earlier episodes, we were kind of highlighting that, you know, if we, a year ago when we were using these tools, I think we all had like a healthy dose of skepticism and we should still maintain
that healthy dose of skepticism, where we saw well, these tools are hallucinating a lot, they're making simple applications. Bret: I spend all my Nirmal Mehta: Like, Bret: Yeah. Nirmal Mehta: the output, had to be shepherded looked at very closely. and I think a healthy skepticism look at that was, you know, these models are really cool and there's cool stuff that's enabled by that. And some of these operations, like with MCP servers, that's awesome as that was starting to roll out. But it's not gonna replace this whole entire function or, you
know, the quality is not there yet. and if the last time you used any of these AI tools was, let's say six months ago or older, Then you might not realize something that happened in October and, in the last six months or however it's been since October. and that is these new models that came out, especially, and we heard this from, I can't even count how many people told me.
this was like announced at the Pragmatic Summit as if it was like just an obvious paradigm shift. But when Opus 4.6 from Claude came out from Anthropic, Claude, Opus 4.6 came out and sonnet 4.6, and some of these other models too. Gemini 3.1, pro 5.3 Codex, Bret: right. Nirmal Mehta: like basically the, Bret: this year, but yeah. Nirmal Mehta: let's just say like the latest, Cohort of models that have been released. Bret: while you were, let me recap for you.
While we were at KubeCon, We had multiple state of the arts. We're gonna call it the soda models, frontier models, whatever. Released like within the same week. One of them was Opus 4.5, right? We had Sonnet, I mean we were on Sonnet three five a year ago Nirmal Mehta: right. Bret: Sonnet and Opus 4.6. Many people. Yeah, like you, so many people have told me are game changer. It is so much better. The hallucinations are way less. We now have the AI guiding us through gaps in the context.
I just demoed a couple hours ago on the live stream how if I tried to build an ECS cluster and it's walking me through, answering the questions without me asking it's way better, Nirmal Mehta: Yes, Bret: much better than six months ago. Nirmal Mehta: correct. Bret: I think everyone who's done this or tried this is agreeing. I think like you're saying, everyone who's still using the cheaper models or maybe not the Claude models, I mean, it's not that like Claude's got. is the cream of the crop right now. It's everyone's favorite. That's a combination of the Claude code CLI and in
combination with the Opus models. But can still do this on Gemini Pro 3.1. You can still do this on code, open AI, you know, chat g or GPT 5.3. Now, Codex, if you specifically look up the Codex, but if you're using either one of those three, I just know that I have the most experience with Opus, so I can speak to that. But what you're saying is both of us at the same time in different groups of people across Nirmal Mehta: Correct. Bret: the industry, within two months or even just a month of that November release, everyone I know who was what I would call blue pill,
meaning we're not really convinced AI is gonna do that much for us. and maybe you can get the function right, but that's the best you're gonna do. I'm probably gonna have to change something in that function. You're probably gonna make unnecessary lines of code that you don't need. And these are all still true, but they're way less common. And now suddenly that combined with the lead release of skills, which also happened in the fall in October, Claude announced the idea of skills that has taken on a whole another level of hype because it's truly easy to use and I am actively creating new skills every week for the last two months.
So we are now in a world suddenly where six months ago I wasn't sure where AI fit in my world. And now every conversation for everything I do in DevOps starts with AI. there's nothing I'm doing anymore where I'm not aligned with an AI assistant who is helping me walk through my design or my GitHub workflow creation or, you know, whether I'm making a landing page for something or whether I'm deciding whether,
The CVEs I wanna patch are even necessary. there's just so many different things. I'm doing AI first now and I think in the last 24 hours, which I've spent probably over 10 hours in the last 48 hours talking and waiting on an AI. And it has only gotten I think, one, maybe two things wrong. And they weren't even really wrong. They just work the way I wanted 'em to. And that is across dozens and dozens of prompts across multiple projects GitHub Actions, on web
Nirmal Mehta: Yeah. Bret: GoLang tooling for my DevOps. Like I'm just doing so much stuff all at the same time. So you and I now have to bring that, like to me the goal is to bring that experience to the enterprise team, right? Nirmal Mehta: Yeah. Bret: the possibilities. Nirmal Mehta: I just wanna, highlight this and make a very exclamation point for our listeners. first of all, thank you all for the folks that have supported us and, shouted out that they enjoy the show. They trust us too give them some to ground what we've been seeing.
And so I think there's two things here as takeaways. One is if you've tried these tools and models in the past, like six months ago or older and you haven't touched them again since then, I highly recommend you at least try some of the latest SOTA models, and maintain. Another thing that we talked about early on in the podcast episodes is also maintain a healthy skepticism.
Everyone's use cases and what they're doing is very different. they still are LLM models. They still act in the LM model way. and yes, the quality in terms of output that we're seeing for whatever we're working on is increased dramatically. as engineers, as practitioners, as operators of production systems maintain a healthy skepticism, which means understanding how to use these tools, understand where, their limits, their capabilities.
I would highly recommend against just dismissing these tools completely. Bret: the number of people that we respect in the industry have in the last three months gone what I called AI red pilled. This is a Matrix, this is back to the ultimate choice of Morpheus and Neo in the Matrix where he offered him two pills, the red pill and the blue pill. The blue pill meant you'd wake up at home, nothing would've changed you, you know, the reality hasn't been altered, and you can just live your life
pretending that nothing ever happened. That's the skeptics, right? The number of people that, but have, are no longer taking the blue pill and are now on the red pill, which is, let's see far, how far the rabbit hole goes. And let's try this thing everywhere. We can try it. You can still have skepticism in trying a new tool, right? Nirmal Mehta: Yeah. Bret: skeptics when we first started Docker. We were like, this is really that good. and then we saw people that would try to use too much Docker, right? Like you use Docker everywhere. Like we thought it was gonna be the replacement for brew on our local
machines and we were gonna run every local CLI tool in a Docker container. 'cause it's convenient and it's isolated Nirmal Mehta: I mean, we we're kind of going there Bret: We didn't, yeah, we Nirmal Mehta: with agent tools. Bret: far. Nirmal Mehta: Yeah, that's true. Bret: I thought containers were gonna be like I thought Docker files were gonna be our build documentation. Like it was gonna be everything for CI inside a Dockerfile. Nirmal Mehta: True. Bret: were stages and Docker files. Nirmal Mehta: Yeah. Bret: thought I was totally red pilled. Obviously it's always somewhere in the middle, Nirmal Mehta: Yeah. Bret: I like that you wanna be skeptical, but I am not issuing the same skepticism. Nirmal Mehta: I think,
Bret: mean, Nirmal Mehta: think what I mean by skepticism here is have a systems engineering approach to these new tools, right? don't just read a blog post and. Consume the hype and say, oh, it's like a hundred x productivity gain across the board, but really start to use these tools. I think the broader message here is, don't ignore it. Get good at using these tools Bret: Right. Nirmal Mehta: think about how these tools can be used in your day to
day, in the automation of the systems and, orchestration and automation and documentation that you might be responsible for in your job or your role. And also think about it as like an artist would, a canvas, a brush paint where. You could use these tools to explore lots of innovation and creativity as well. but skepticism in the sense that use it, see how it's being used.
Don't just take like LeVar Burton said, don't take just our word for it. Bret: I don't think anybody, we're not even using the words vibe coding anymore. those words are a year old and we now understand the nuance and it's sort of a meme at this point, but no one's saying Vibe code your DevOps. Right. Nirmal Mehta: Correct. Yes. Bret: That is why I'm making courses. That's why I've been working for a year now to build this DevOps world in my community and in my content, because this is gonna take years and years. It's, this is every one of these big waves like Docker, like Kubernetes.
Nirmal Mehta: Yeah. Bret: It's at least 10 years before ma. The majority of companies have understood it and adopted it, but I think there's three main areas for people to focus on in the world of DevOps. The first area that everybody should be doing right now, regardless of your skillset, regardless of whether you're allowed to at work, just do it at home, is to try, even if you can only get the free models. GitHub free account provides free copilot, which provides free LLMs for you to use with all of the copilot tools. That's web.
VS Code. That's command line TUIs, like you can use all that stuff. There's free Claude, there's free, GPT, there's free Gemini. You can use these today. But you, if you can just get a $20 plan with any of those things, whether it's GitHub copilot, whether it's Claude, whether it's Codex. just spend $20, even if you just spend it for a month to do some learning, if you can't get your company to pay for it. think anyone that I would sit down with and help them through the basics of getting started would see really fast benefits to their
personal workflow for DevOps. 'cause when we think of DevOps, we think I'm largely writing YAML or shell scripts, or maybe if I'm fancy I'm doing Golang or Nirmal Mehta: Right. Bret: I'm doing some automation for my GitHub. I'm probably spending a lot of time in Git. I need to do git commits and GitHub things, or GitLab things. And almost all of that can immediately be benefited just by having, Claude Code installed on your machine or the chat GPT desktop is a fantastic tool, and you don't have to necessarily be a developer to do all of this to see
the benefits, but it's a study buddy. It's a work buddy. It's there and it's reliable. The better, the higher end and more expensive models you can find more likely will give you better results, Nirmal Mehta: Yes. Bret: seek on open, open weighted models or closed source models. But I can tell you that for the last three months, everyone I know and across dozens and dozens of people, big or small, they're all talking about Opus as the model to use. Some of them still use Gemini and still use GPT 5.3. Those still work, but they're all in love with Opus.
That's the first wave is like get used to local tooling. Nirmal Mehta: Mm-hmm. Bret: push yourself to try using AI as a buddy to help you anytime you need to create some content. Content meaning YAML or whatever. next phase is getting AI to help you automate your CI. Nirmal Mehta: Yes. Bret: we are still very early days, but in the fall of last year GitHub announced a bunch of different things, especially at GitHub Universe. I'm sure Amazon also launched months of different things on
Bedrock, but I unfortunately don't pay attention to everyone. But I can tell you that on the GitHub side, and maybe you can, compare this to what's happening on Bedrock, and other functionality AWS, but like at GitHub. They launched, not only keeping track with all the models. 'cause if you didn't know GitHub copilot models is kind of like open router where you can use everything from Opus to Gemini to DeepSeek, to GPT, to Qwen, to, tiny Little Pie and tiny little, there's dozens of models that you use
all through GitHub, and you only need a GitHub account to get started there. You don't even need to spend money. But if you pay for the GitHub copilot, I think it starts at like $17 a month, you can get access to all the bigger models, and I use that every month. And I can easily go days or weeks in the month on just a $20 plan without hitting limits, and you can start to experiment something that teased us with in the fall, but officially launched. I think actually last month that they're calling Agentic Workflows. Luckily, Nirmal and I were on point with the name of our podcast, Agentic
DevOps Podcast because they're calling this functionality Agentic Workflows. And it's a little bit different than just throwing an LLM prompt in your workflows or pipelines. It's using AI to create workflow. So you create a markdown file with some front matter. You give it the front matter as sort of like the deterministic stuff, and then you write a prompt of what you would like it to do. And then the AI through GitHub Copilot will create the workflow for you and run it.
This is one of their ideas for how we might really complex or even workflows that have an undetermined, non-deterministic nature to them. A good example might be triaging. Why your GitHub Actions failed. So it has to go read the logs, find the failure, then go look up what the failure might be. Or maybe it just knows built into the model, and then it might suggest on the existing PR it might open a new PR based on what you want it to do, and it will
start to help automate fixing problems essentially that before if you're trying to just create a very deterministic way without a Code Rabbit or a GitHub Copilot, those might be very hard to do, right? There would've to be some grips involved, some set and ox, like you would definitely have to have a very long and thorough app. it would, that's what it probably would turn into an action that's almost like its own app. And we've had those, we've seen those before, but now you can start to create these custom ones yourself. So that's stage two. I think stage three in the last stage is where have an always running AI
that's able to respond to events. an operator would. Nirmal Mehta: Right. Bret: the most mature area. So in terms of the maturity model that I'm building out for my courses and for my community, it's starting with the local tools. You and I saw that like the individual developers are very productive on their local machine with these tools. as you get really comfortable with that stuff locally and you figure out what YAML and TOML does this thing write well, and where do I need to learn how to hold its hand a little bit more? You're gonna learn all that. And then as you get better at using what we call now the agent harness,
which is just really the UI that you're using with your agents, but I'm using in the open code harness. Others, might be using it like Claude Code or Codex or a different harness, Nirmal Mehta: Yep. Shout out to Kiro. Bret: shout out to Kiro. Yeah. another harness, like there's a lot of harnesses and everyone's got their favorite and they all sort of are merging around the same ideas of letting the agent continue to go. Longer and longer, like we're getting better at the AI's, running for longer periods of time, doing multiple steps without hallucinating. That's getting better. Nirmal Mehta: Yes.
Bret: time, we've invented this idea of skills that really allow you to create a SOP, a standard operating procedure for each type of work you do. And I just talked in the livestream today about, as you get better with your local tooling, you have your own work that you typically do over and over. Most of us don't create YAML and TOML, out of thin air, right? If we're doing DevOps, we're usually copy and pasting a lot. We have standards in the team repo that we set up with standards in them that we use our official templates from, right?
We've all got different processes for how we know that this particular cloud formation has to be, that these adhere to these standards, pass these lins, like you've got all that. It's probably in your confluence or in your GitHub repo itself, you can very quickly turn that into a skill. That skill is just markdown. And it's like you treat it like you're teaching a junior engineer to automate or to do some part of work that you don't wanna do yourself. And so I have had great success in the last month the standard operating
procedures that I have for how I create a reusable workflow in GitHub Actions or how I would go about creating some sort of terraform. I know what I care about. I'm particular about certain things. I have certain standards those maybe are in my head. We call that tribal knowledge. I'm a tribe of one, so I have tribal knowledge to myself. But you're gonna put that in a markdown and I just call it a SOP now. Like the skills are SOPs. I had SOPs before in my notion with my team that we are automating through skills and in AI today. And I think that.
be a bigger portion of my courses that I anticipated. ' Cause once you can get good to that, you're onto something because what you've done is you've documented a non-deterministic workflow an AI that runs locally on your machine, you're only one step away from that running in a GitHub action. And that's what GitHub calls Continuous AI, which is AI running in your automation, doing various things in a non-deterministic way that you couldn't
otherwise do with the previous generation of what we're calling deterministic workflows, which is programmatic if then statements, L statements. Like Nirmal Mehta: based. Bret: Yeah. So I need to have a diagram for this. I don't. but this is, I think where we're all going. And I think that if you're not at the beginning, at least of this series of steps and maturity model. I think now is the time, baby. Like we're a year into gentech DevOps and today is the day. Nirmal Mehta: Yes, totally. And I think it's somewhat optional to do.
I don't think it's gonna be optional for, for our jobs in the near future. And I think it's not gonna be optional. I think we're gonna need these tools to support the proliferation of CI/CD workflows that are gonna be needed to deploy, And maintain all of these new applications that are being built with these Agentic tools. I want to kind of highlight something here.
What we saw at these events in the last month were developers accelerating the, their application development lifecycle, the inner loop, and then they weren't making decisions on how to deploy these applications. Claude Code was doing that. the agents were trying to figure out, they let their agents figure out how to deploy this. And so those SOPs, the concept of SOPs and skills are gonna be super important because, as operators, we're gonna be responsible for creating
the platforms and the guardrails. That are the landing zones for these applications and also the system that helps operate all of these new pipelines and workflows that have to be built to deal with a, you know, what I call the tsunami of applications that are coming to you, to all of us as operators and Bret: gonna be creating like this. Nirmal Mehta: I think we're gonna be create, Bret: Yeah. Nirmal Mehta: Yeah, I think that's true. I think the word DevOps is gonna be.
Even more DevOpsy. like we're really gonna be developer operators, with respect to building applications that are doing all of those operational tasks, for us, which then leads to the next thing, and I know this is maybe a little bit skipping in our outline today, but that means that we're gonna have a lot of agents running and doing things on our behalf, running kind of in parallel. And so, if you've been kind of just seeing what's going on in the last couple months,
I think the term agent orchestration is a term to look out for and I think as operators, as platform engineers, as SREs, we're gonna have to be able to answer how. And where, and how all these agents are gonna run in a secure environment, with all of the, well architected principles around it. And, that's another area of tooling that is just coming out right,
is, and I think, one can make the case that something like OpenClaw is like agent orchestration layer. I think Kubernetes is very well placed to be an agent orchestration layer. And, I think all of our cloud native Kubernetes knowledge and containerization knowledge, is not obsolete. It's gonna be very well placed to be the harness and the orchestration for all of these agents as they're running, doing all of these tasks, both on the
developer side and on the operations side, and also your business agents. We haven't even talked about that. But I think another thing that we saw is non-technical users starting to use these tools to build business applications that aren't devs. Right? They don't know anything about CI/CD. They don't know anything about infrastructure. They don't know anything about code and they're gonna start building apps. And, if you thought Shadow IT was a big problem back when cloud was new, oh boy.
Shadow IT with gen AI is like, that's a whole brave new world that we're entering into that. Bret: which really just goes to the fact that we're gonna need more automation. There's a great graphic from our friends at Geo Coio. this is the dev loop, or I guess it's really the inner plus outer loop. This one, you saw this at the retreat. so this is by Michelle Hanson, of Geo Coio. And it's a very simple diagram. in the first one, it's traditional projects before AI. And it's basically a timeline. And one fifth of this timeline is scoping before development, and then
three fifths of the timeline is coding. And then one fifth at the end is qa. And what they're describing as with AI, maybe you can get it done in 60% of the time as before. So, so you're saving 40% of your time overall, but the coding is highly compressed to less than one fifth of the time. This is just approximation. these aren't exactly like Nirmal Mehta: Right, Bret: metrics or studies or anything like that, but the scoping is doubled and the QA is almost doubled. And so the two sides, the planning at the beginning.
Which we now have plan modes in our agents, Nirmal Mehta: right. Bret: now have these concepts of plan agents and build agents. Nirmal Mehta: And specs Bret: helping you create the spec. Yeah. Nirmal Mehta: yeah. Bret: plan with all the tasking. And then at the end, you're gonna need a longer qa. QA to me means I'm gonna need more automation. I'm gonna need code ql, I'm gonna need way more AI and automation running against every commit because I'm gonna suddenly get a bunch of projects thrown at me that weren't as well, vetted by humans as they were in the past.
Not if AI is making lines of code, that means that humans may have at best read the code at best, but they certainly didn't write every line of code, so they didn't rewrite it. They didn't rethink about every single line. They didn't toil over every line, which maybe in the long run is actually a good thing. But what that really means for me as a CI person is that I'm probably gonna be more responsible, and I've had already two. Now, in the last month, two different teams tell me They're having to push back as DevOps engineers to the developers
because the developers are shipping more vibe coded product that they're having more difficulty because their testing isn't catching the problems that the AIs are creating. And so these apps are getting shipped to production. They fail in production, and then the developers go, it works on my machine. and this is the age old problem, but it's happening again. Because we're shipping more lines of code in a commit. We're shipping more lines in a pull request, and not as many eyeballs have studied every single line.
the leading edge teams in this field of AI right now, the ones that are pushing the limit, aren't looking at the code anymore. Like they're shipping production code. That's only, that's written by AI, and then reviewed by AI. So, and you can have three different AIs review it, right? you can have Nirmal Mehta: Yeah. Bret: my friend Aaron Francis, created a new tool called Counselors, which does that exact thing. you give it a problem. it has three or more AIs it will go and ask, and then it will summarize the differences and what they agreed on and what they disagreed on in a summary for you.
And It's literally called the counselors, And we don't have that specifically for DevOps yet, but I can see that coming. Where we're going to have a council review, the PR, because maybe this month opus is a little bit more loosey goosey, like Opus likes the vibe coding this month in the latest version. And maybe Gemini is more picky by default, right? this reminds me of the nineties and two thousands with antivirus software where we had, the best antivirus teams that, like I ran an antivirus team for a, enterprise of 7,000 mailboxes.
we never could trust just McAfee or, you know, or just one scanner. So we had the scanner of scanners. We had a single product that had like eight or nine. It had Kaspersky it had all these different scanners you had to be able to get an email through all of them it would get to the inbox Nirmal Mehta: I think so. Yep. Bret: we had very opinionated, and the same is true the CVE scanners today, like some teams. Scan with two different CV scanners. Some teams, have their favorite. But anyway, this image that I'm describing to the audio audience.
the second thing, do you have any thoughts about Nirmal Mehta: So Bret: tell you Nirmal Mehta: yeah, I think I saw this quote we ship faster but break more. So that's like the era that we're in right now, right? Like we're shipping faster, but everything's breaking more than it used to. Bret: yeah. We've got a, and that leans on the DevOps engine, Nirmal Mehta: Correct? Yes. Bret: automation. That's, and there's so many talks over the last year that I could quote, Nirmal Mehta: And. Bret: Han's doctor at the Gradle Summit, I think the CEO of Gradle, he talked about this.
He's basically saying a wave of problems coming to DevOps and SRE and Ops from all of this. And we better prepare because it's come. Nirmal Mehta: Yes, Bret: to your team, it's coming soon. Nirmal Mehta: And the other thing here, just going like highlighting that diagram, scope and Q&A scope is documentation, it's context. And, we haven't had this conversation yet, but I think in the new architecture there's going to be a context layer.
there's gonna be agent orchestration, but the's a context layer there As operators are gonna be responsible for main, like creating, run, running, maintaining, because when no one, no human's looking at the code, the code is still important because it is the application, but it's, importance is goes down a little bit compared to the specs, the requirements, the context, your business documentation and your Jira tickets.
And like the context and the intent of your developers or your business users is more important than the line of code. Bret: Mm-hmm. Nirmal Mehta: And I think that was another theme that came out of our conversations the last month is that, Having systems that maintain and version control that context are going to be part of our architectures moving forward as well.
Bret: Yeah. Nirmal Mehta: And, what that looks like. Not sure. I think it's gonna be an amalgamation of a lot of things, but this also means that with skills and these operational agents, these i AI ops agents on the SRE side, that's still true there. It's the context. You know, it's your runbooks, it's your SOPs, it's your troubleshooting steps, which your in incident response, which are your product documentation. If you're sitting there listening to this and, you've been fighting your management about having, story points, in that sprint to update your docs
and like never getting enough time, This is your hour, this is your time. Because organizations that will be successful with these tools are the ones that have automation in place or building toward that have good documentation practices, have good, operational practices and organizations that have been deferring that for another time in the future. That tech and like have not paid that technical debt, that you're gonna be behind.
Just straight up. You're gonna be behind, you're not gonna be able to leverage these tools because that's the main interaction with these tools. Now. it's not like programming a line of code. it's giving them context and intent. Bret: you know, when we talk about these things like the skills and the agent files and the rules and the commands, I was talking on the stream today about the difference between skills and commands. And I think skills, I mean, commands are probably mostly dead. I think skills are gonna win and like that, I think that model is people are really understand liking that model and it's working well for people.
But we're essentially just recreating the software lifecycle we had documentation in wikis and doc systems, and we had standard operating procedures in another system where they were very meticulous and they had lots of screenshots and like, this is how the humans get things done. These are our standards. And we would have, you know, mermaid graphics or various other tools, you know, Visio or whatever you might have. For diagramming workflows or diagramming system designs or network designs. Like these are all things we would do in planning.
And then, then after we'd ship something, we'd realize, especially like when you hire new people, you'd realize, oh man, like we're, there's a gap in our knowledge for them. Nirmal Mehta: Yeah. Bret: I realize we missed this process. Let's go create that sop. these are all things that the individual or small team developers are figuring out how to do for the AI. 'cause it turns out, if the AI has all of that, he can perform or she, or the it, I guess it's an it. The, it will now look at all that stuff. And the simplest way for us to do that today is to just put that in the repo.
The problem for like us DevOps people is that we don't. In very few place teams that I've worked with is all of their DevOps knowledge. In the repo it's, I mean Nirmal Mehta: Right. It's all over. Bret: a lot of TOML, yeah Nirmal Mehta: Yeah. Bret: but they have documentation systems. They have Slack, they have tribal knowledge. They have all these different various places. They have some stuff that's basically in the dashboards, like you imply knowledge through a dashboard layout and what things are important and which metrics we care about Nirmal Mehta: Right. Bret: these are all over the place, but the reason it's all in the AI repos today or the AI enabled code repos today.
It's not because I think that's where it's all gonna be in five years. It's because that's the quickest way we can figure out how to dump text into the AI context so that the AI is way smarter. And we're eventually gonna figure out whether it's MCP that connects all these things together. Like we might very quickly end up moving past markdown in directories to, Hey, if you just point this to your MCP to Confluence, then you can just point it to the different, you know, URLs of all the DevOps documentation and you don't need the agent file, Nirmal Mehta: Runtime.
Bret: Yeah. we're gonna figure all that out. These are, those are like the standards and the practices and the conventions, I think we're all gonna figure out. But like right now, since we're in the bleeding edge, it just happens to all be markdown with SMO front matter and a repo. So go check that stuff out. the last thing I wanna talk about before we wrap up, 'cause I know we're hitting our limit, but this is the story I tell to people that I, this is my vision now. Like I didn't really have this vision nailed down a year ago. I think. think everything I was saying, I have to go back and listen to our first episode 'cause I feel like everything we're predicting.
Has come to pass. Like it's coming faster than I thought. Thought like, I Nirmal Mehta: Yes, Bret: it Nirmal Mehta: that's for sure. Bret: before we had reliable models for DevOps work. this is not about replacing our jobs. Like everyone's concerned about it. I'm concerned about it, especially as a content creator because course sales are down I wouldn't say no one. People are buying courses less and they certainly don't need courses that teach them the manual anymore, which is why my whole course vibe is completely shifted to advisory and best practices and architecture type stuff.
Not, Hey, this is how you need to learn every single command. 'cause you're not probably gonna type the commands here very much longer. I started 30 years ago and this slide I'm showing on screen for those audio listeners is about the major shifts that about once a decade, although these are increasing now to about two a decade of major shifts in it I see as an operator and a DevOps engineer that have been happening my entire career and are gonna continue to happen. And it just so happens that managing agents, which are a bunch of LLMs
is just the next phase of it. And in the nineties and this graphic, what it's supposed to represent is at the top is in the nineties I got to be a part of the mainframe to PC wave where I was literally like shutting down a mainframe or mini frames or Unix servers that were, stuff Nirmal Mehta: X. Bret: to pc. Yeah, shifting to PCs that didn't even have mice yet. Like they were literally dos with word Perfect. Right. And then we eventually got mice and we got windows, and then we got Windows 95 and then. Like we started to do distributed computing and that enabled me to go from
managing one large mainframe, which I went to voc, two years of vocational school on how to manage Honeywell and HP and Solaris and these kind of things, right? but I was only able to manage a couple of these things. We had like four dudes to manage five mainframes and many frames, and that was it, right? So it was like a one-to-one. And then in, in the early two thousands, we invented the technology as VMs, virtual machines. And I can remember those days very distinctly because everyone told me at the, my enterprise, I was working at a large city, half a million people, and.
Man managing a city of a half million people. And I remember Nirmal Mehta: I. Bret: the meeting where we were just trying to describe how VMs work and none of the system engineers thought it was a good idea. Everyone thought it was gonna wreck, shop and crash things, and it would be horrible. And we were having to give analogies, like talking about like the blade that goes inside the mainframe is actually a different compute that you can shard it off. And this is kind of like that, but in software. And everyone's like, no, it's a horrible idea. We're running a kernel and a kernel. Well anyway, VMs are the standard. They have been for decades, and it enabled us to easily manage 10 or
even a hundred machines ourself. Then you fast forward to the cloud now system engineers and DevOps people can manage hundreds of servers themselves. They, that's the dawn Nirmal Mehta: say thousands at this point, Bret: thousands. Well, we had to create new tools for that, right? That when we Nirmal Mehta: right? Bret: that in 2010, we didn't have Nirmal Mehta: Yes. Bret: tools. Nirmal Mehta: Right. Bret: then Ansible shows up, then Terraform shows up, and now we're able to do thousands. Then containers show up and we stop talking about servers. And now one engineer can manage if not thousands, if not tens Nirmal Mehta: Ends of thousands,
Bret: workloads. Nirmal Mehta: correct. Bret: happens in the 2010s to, you know, Kubernetes takes us all the way up to 2020. And then we get things like off shooting from that, like Wasm, serverless, lambda, stuff like that. But that is not replacing engineers. None of those things in my job replaced anyone. I literally took the same two guys that were managing hardware installations in the data center. They were so worried about their jobs in 2005, and I gave them jobs managing the VMs. We had to retrain them, retool 'em, but now they had VMs to manage
instead of physical hardware. And the same thing happened. We went to the cloud. We had to take everybody that was managing SSHS to a bunch of Linux servers in a closet. And teach them cloud tooling. And now we have better cloud tooling so we can manage even more. We had to teach people Docker so they could manage thousands of workloads instead of thousands of servers. And I think that AI today, if we keep reiterating this over and over, that I need to change this slide now that we're just gonna be managing a series of agents, gonna be directing agents all day long. I'm already doing this I'm like, I need you to create this GitHub action workflow.
Please use my skill that outlines my favorite things that I always want in my GitHub action workflows. Please help me meet these goals and ask me any questions along the way that you need answered. Turns out the new opus models do that really well. It comes back and asks you questions with a nice little menu system that you get to choose what it recommends versus what you think or do I wanna opt in like to a hand filled, crafted answer to the question? It interviews me just like a junior engineer would when they go, well, you didn't teach me. How you expect the security to be in this YAML file.
And I would realize, oh yes, let me teach you young Padawan. That's the same thing I'm doing to AI right now. And I'm able to do that thing is going off and spinning up a new workload that is building a new, workflow for me. And then I hop over to a new terminal where I'm SSH into a server, and some people are going crazy and actually installing these harnesses, like on running servers to troubleshoot the server. I wouldn't go that far, but maybe I could use my local one to SSH into a server and it could send SSH commands one at a time maybe. I feel like probably gonna be a day where we have a tiny little
agent harness that's like a single binary that doesn't touch anything. It has no right access. And we can maybe throw that on a server temporarily. Maybe it runs in a container and then we pull it off when we're done and no one ever knew. essentially there to help us troubleshoot. I mean, you have probably heard of the net shoot Docker container that our friend Nikolai has created Nirmal Mehta: Yes, Bret: now. Well, that's gonna get replaced with an AI that has all those tools built in and just goes and does the Nirmal Mehta: correct. Bret: us. Right? Nirmal Mehta: Correct. Bret: it's all this really means is I'm managing more infrastructure, more
automation that I have to do because there's more code being shipped. More and more people making code now because now we're talking, we're hearing about product managers. They're able to code again because they're able to actually make commits that they wouldn't otherwise have time to make because they can do it in a fraction of the time. this is just gonna keep happening. So I'm gonna just help everybody feel comfortable that your job isn't going away. It's just changing maybe faster than it would in a normal 10 year arc. Maybe this is gonna be like a five year arc
Nirmal Mehta: Yeah. I mean, we're already two years in. Bret: using AI. Yeah, Nirmal Mehta: Yeah. Bret: I Nirmal Mehta: I think you're right. This is like, this is a five year arc. Bret: Yeah, Nirmal Mehta: so that was well put. I think on an earlier episode, I reiterated, you know, this won't replace you, but someone who knows how to use these tools will Bret: Yeah. Nirmal Mehta: if you need some impetus when you're sitting there trying to figure out what question to answer for yourself, with these tools, think about how would
I use these tools to do some small aspect of my current job day to day, but also think about how would I use these tools to operate an environment with instead of a hundred containers or a thousand containers, maybe 10,000 containers or a hundred thousand application containers Bret: Yeah. Nirmal Mehta: How would I use these agents to do that? And maybe that's a good question, for exploring these tools is
like getting to that answer. I would use these tools in this way, to operate systems at that scale because I think you're right. we've seen this trend from mainframe to containers and serverless and cloud, that one person's scope of responsibility has just expanded, but the surface area of applications has also expanded. And so there is a need, and it won't eliminate your job. It's just the person that knows how to use these tools will be doing your job.
Bret: Right, right. Nirmal Mehta: So hopefully at the end of this we're almost close to the end of our context window for this episode. hopefully we've convinced you that you could still maintain that skepticism about, the output of LLMs and these tools, but. Please do yourself a favor and start using 'em and start getting used to them and start understanding how to use these tools effectively. 'cause it's just like any other skill. it's, you know, you have to learn, you have to exercise that muscle and you have to understand what it can and can't do.
with that, I mean, we got halfway through our list today, so tune in again for another episode. We've got so much more, Bret: So what you Nirmal Mehta: context to explore. Bret: coming, want me to Nirmal Mehta: Yeah, let's, Bret: that are coming up. Nirmal Mehta: yeah, let's do it. Bret: D, we're gonna have Solomon, hes at Dagger back on the call. We're gonna have, the co-founders of Mineral, which is a GitHub Actions AI. Troubleshooter that essentially I've been using for now a month that it just stares at my GitHub all day long its own running 24 7, looking for ways to help.
It doesn't do code commits in the classic way. that's a different thing. It's not trying to like build my apps. What it's doing is looking for failed GitHub workflows or, CVEs that Abott has suggested that I need to implement through a PR and it's reminding me that those aren't there or it's recommending strategies for improving the speed of my GitHub workflows. it's highlighting whether the CVEs are necessary to get fixed in a particular depend bot or renovate PR.
It's, there is so much stuff that it's doing, but it's essentially like. I look at it as like the janitor, like the custodian of my repos, and it solves a very specific problem set that I'm excited to have them on to talk about because I'm Nirmal Mehta: Sounds awesome. Bret: I feel like I'm currently a free customer, but it's providing use to me as an individual. I can only imagine how helpful it's gonna be for a team. it's kinda like an automated Jira. Like it makes the tickets, the fix for the ticket, and then if you just say approve it fixes it with a PR or whatever it needs to do to fix that thing. Nirmal Mehta: cool.
Bret: And so we've got SRE ones coming up. I'm gonna try to get someone from any shift on, which is an SRE type of tool that's similar to the mineral. we're gonna try to get someone on from the leadership of GitHub on, soon. And so we're gonna actually have someone talking about GitHub Actions that actually controls the teams that are working on that there. we're gonna, I'm working to get someone from GitHub next on, because GitHub next is where a lot of this innovation is happening in the world of what they're calling continuous AI or Agentic workflows, which is all this stuff they're
doing with co-pilot inside of GitHub. and then that's probably only half the list I have that we're currently working on. So we've got more episodes coming. I'm super excited to finally get into some real data around how do I implement a workflow that's not gonna hallucinate or be at risk of a prompt injection or it's gonna leak credentials through MCP. I'm super ready to get into the nerdy details, and that's what I think this year is gonna be all about. Nirmal Mehta: Awesome. And with that, please let your colleagues, your friends, your family, your neighbor, you know, or your mom,
Bret: Agentic DevOps Nirmal Mehta: to subscribe. And like the Agentic DevOps podcast, you can check us out at Agentic DevOps dot fm if that's not the way you found us. please share, Bret: favorite Nirmal Mehta: we'd love to hear from you, if there's a specific topic that you would like us to cover or go. dive deep into, please let us know. We've got a long list and, can't wait to dive into that in this season of the Agentic DevOps podcast. Bret: yeah. Find us both on LinkedIn. he's Nirmal Mehta. I'm Bret Fisher.
We're on LinkedIn. We're on Blue Sky. I'm still on Twitter. You can find us on our discord@discord.gg/devops. All the links are in the show notes for what we talked about today, and we will see you on the next episode of Agentic DevOps. Chow. Bret AI July 2025: Thanks for joining us, and I'll see you in the next episode.
The OpenAI Hugging Face hack. The Anthropic testing failures. Listen to the security and ops experts, not lab researchers. I walk you through the production failures that allowed OpenAI’s models to access the internet,...
Luke Hinds, co-founder of nolabs Inc. and the nono project, joins me to dig into AI agent sandboxing with the nono CLI. It's been my go-to sandboxing tool for local agent use, but I know there are more features I should...
My friend Mike Rollins joins me for a chat about what it’s like leading a software engineering team that doesn’t read code anymore. Click here to watch a video of this episode. Thanks to SpeechifyAI for sponsoring this...
A deep dive into GitHub’s Agentic Workflows feature, with Don Syme of GitHub Next & Peli de Halleux of Microsoft Research. Check out the video podcast version here: https://youtu.be/zmM8VISTOwo 😇 My new GitHub Security...
Bret’s joined by the Anyshift.io co-founders, Roxane Fischer and Stephane Jourdan to discuss how an always-on SRE agent can help you proactively find risks and avoid incidents. 😇 My new GitHub Security workshop has...