# Strategic AI Agent Deployment and Workflow Optimization

**Podcast:** Dev Interrupted
**Published:** 2026-05-22

## Transcript

Hey, Andrew.
So are you giving your agent a phone number yet?
I heard this is like all the rage all of a sudden.
Well, you know, a lot of parents won't even give their kids a phone until they're 13.
So why would I give my agent a phone when it's like three months old?
Exactly right.
That's terrifying.
What about you?
Would you equip your agent with a 10-digit cell phone number?
Well, you know, I didn't even have to think about it for like even a second.
I was like, obviously the answer is no.
But then my second thought was, oh, we're now at the future where AI is just going to be spam dialing me through OpenClaw and PolyReach is this thing that we're talking about right now where you can apparently give your agent a real number to make phone calls, which just seems like a nightmare to me.
It's really interesting, the idea of pairing your agent with.
a phone number, a PolyReach's take on giving your agent more of a presence in our reality is they're not alone in taking that stake and trying to bring agents closer into the world we live in.
Because I've seen versions of this kind of tool that give your agent like a bonafide email inbox somewhere or even a bank account.
And that's not even to say for all of the agent activity that happens on places like on blockchains as well.
So you have a lot of people trying to equip them.
interact with the economics of our world.
But personally, I'm going to save the phone number for later.
Maybe we can just stick to the we can't we just stick to the telegram chat?
Like, why do we have to communicate this?
Yeah, yeah, for sure.
I will say, though, that I've been in places where I don't speak the language and I would like to make a reservation for like a certain time, you know, and being able to like have a service that could call on my behalf and like handle that for me.
But of course, I would not trust my open claw agent connected to a phone number to figure this one out.
Like, I'd be too paranoid that I'd get something wrong.
There is something there about the idea of turning the agent into your phone assistant or your own, you know, personal assistant.
Because as we know, like phone menus and calling in somewhere to talk to a human, they're with their.
obstructed with all of these walls purposefully meant to slow us down, is purposefully designed to be a bad UX.
Because frankly, most companies are optimizing for reducing those number of calls and handling them before they get to an agent.
So there's so much incentive to actually keep you from doing that.
By putting the power of an agentic phone assistant in your fingertips, it kind of puts more power in the consumer's hands because they can send that agent, that assistant down, you know, the IVR menu or the wormhole to eventually talk with somebody.
It can sit on the weight music and listen to, you know, the metal music from the 80s.
And you can just do something else and it'll let you know when it's ready.
But this actually isn't anything.
brand new by itself.
The concept I'm describing was actually some of the first things that we saw from Google on like the Google Pixel system, I think even years ago of it being able to call in and handle like a wait time on your behalf.
So you're just going to see more and more of these kinds of technologies evolve.
And in this case, maybe being paired with an autonomous agent, you started to get some really interesting use cases.
Well, before we spoil some of the news today, we'll get into Google.
Welcome to the Friday Deploy brought to you by Linear B.
I'm your host, Ben Lloyd Pearson.
And I'm your host, Andrew Ziegler.
And speaking of Google, this week we are recapping all the news that caught our eye from Google I.O., a Vibe coding research project that could use your help, how to distill yourself with AI, and an open source Obsidian alternative.
That just seems like a personal attack at me.
But before we get into the personal attacks, we of course have to cover all of the announcements that were made at Google I.O.
this week.
But the biggest one probably being the announcement of Gemini Spark, which people are out there comparing it to OpenClaw or it's like Google's version of it or whatever.
I mean, in a sense, maybe there are some parallels, but it's basically Google's always on AI agent that runs across their suite of cloud tools.
You know, it can write emails, it can manage your calendar, monitor certain types of things like expenses, and integrates across, you know, all the Google suite, some third party tools as well.
And it really is just designed to help you like work across your, your entire ecosystem of Google products.
And, you know, it just seems like this is a thing a lot of companies are doing right now this type of like agents that operates on their data.
And I get why a company, these companies release or releasing these types of agent frameworks for their products.
But I don't know, I kind of feel like they're missing the point sometimes with how AI can actually help automate our lives.
But I don't know.
What do you think, Angie, before I get into that?
I agree with you.
I think that we're all trying and experimenting with surface areas that maybe get a little too big and unwieldy for the specialized kinds of surfaces that are underneath.
So you start to get this glom of like everything assistance that all are just kind of stepping on each other's toes.
It can be hard to delineate, you know, what are the actual divides?
And then underneath all of this, you know, inflated purpose for the tool, what are the core primitives that are going to be durable that I can...
build on that I should be thinking of as a net new way of interacting with the tools as opposed to all of the abstract stuff that sits on top.
But that said, it's been a really exciting week for Google I.O.
It was really great for Dev Interrupted to plug into that momentum as well.
We had Matthew McCullough, the VP of Android Development Experiences on the show just earlier this week on Tuesday, talking about some of the stuff that I'm going to cover a little bit further down below.
But I just want to really Take what you said and echo it, that people see something like Gemini Spark and they immediately compare it to OpenClaw.
And I think that speaks to the very sticky mental idea of this autonomous agent in the cloud.
It's very abstract.
It gives us something to cling on to.
But I think there's a lot of opportunity to build new services and tools on top of these kinds of primitives.
Going back to what I was saying, open claw, the reason we keep coming back to it and calling things open claw is because it's a primitive.
And the sooner you identify it and figure out what does that mean for you, the quicker that you can leverage it.
Yeah, I think it doesn't get applied that way a lot of times, though, because I think that, you know, AI really works best when it's applied to constrained situations that have like super clear, clearly defined objectives and requirements.
And, you know, he's like general purpose productivity agents are just they're trying to tackle too many things all at once.
You know, I imagine they probably work well for fairly basic and well-defined use cases.
Like when you look at some of the, you know, the messaging around these companies.
They sort of, they get very specific in how they set expectations.
Like it can capture invoices, you know, versus like actually helping you like really go deep into understanding like your inbox, for example.
You know, and there was also like an example of like creating a writing style guide based on your emails, which you can already actually do if you just spend a little time with, you know, whatever AI tool you prefer.
It's just a little manual right now.
And I really think like why, what I see, you know, companies like Google.
As much as I do get excited about these types of products, where I think they're kind of missing the point, is that, you know, like, let's just look at, like, one of the capabilities within the Google Suite, like a Google Doc.
You know, that was still designed and built for the pre-AI world.
And now we're in this, like, era where everything is getting pushed through AI workflows.
I really need something that allows me to collaborate both directly with humans.
Like, that's still a very important part.
But even...
Equally important to that now is having my agents be a part of that experience.
I need them to plug in and take the feedback that comes into me and the tweaks that we have to make and interprets everything that's going on.
And I really feel like these companies would be a lot better served from a user perspective if they just spent more time modernizing these existing tools to work in the new norm rather than trying to build this agent that operates on the old, like the norms of yesteryear.
And I know we've seen Microsoft in particular, I think, fail pretty spectacularly at this with Copilot.
And I'm not talking about the GitHub Copilot.
I'm talking about the general Copilot that they use across a lot of tools.
There's been a lot of memes about just it kind of going off the rails and not understanding how to actually achieve what it's been asked to do, but still trying it.
So I think if you want to give your agents a lot of autonomy like what Gemini Spark does.
You also need to have really extensive constraints and controls in place to keep it working over a long term.
And I think it's achievable.
Like, I think we're going to get further closer and closer to this over time.
But it just feels a little out of order in terms of like how we should be getting there.
Yeah, but on more fronts than just the model and and how folks are interacting with the software, you know, there's a lot of developments and things coming out of Google I.O.
also around how they're servicing things like in the Android ecosystem.
And that's what we talked about on our episode earlier this week.
with Google, but is really a big focus of the Google I.O.
is the ability in AI Studio to, you know, vibe code or agentically code from a prompt, a full Android app, and the domain ability of creating a...
of an android app has a wide level of distribution uh you know so many devices of our world both embedded and user handheld uh they run on android and these models at the same time are getting smaller and more powerful so you're seeing not only these these durable primitives that we've been talking about on the show but also the you know the what powers then the llms both trending towards this getting smaller, being on the edge, being remote, being in your hand, being on your smartwatch, right?
And so I think that's an interesting thing to watch from Google because there's no one else in this infinite space that owns all of that surface area and having as much opportunity to deliver.
You have Apple, of course, which is going to make bets on being the device of choice where people interact with AI and the real world where we live.
But you can't forget about Android.
It's an incumbent in so many spaces.
And here at Google I.O., they really doubled down on that position.
Yeah, yeah, of course, you know, there were a lot of other announcements that were made at Google I.O.
about, you know, the investments that Google is making into AI.
And one of them was, you know, a little bit of an inside peek into the results of all of our token maxing.
So they announced that they are processing 3.2 quadrillion tokens per month, which is up from.
about 9.7 trillion two years ago.
So just an immense, I mean, this is the norm for Google.
They just process immense amounts of data.
And they have lots of plans to just invest into AI as well over the next year.
Some of the product announcements that came out, there's better improvements to their SynthID watermarking system that specifically to make it easier to find or identify AI generated images, which is pretty interesting.
But then, you know, what was probably the most surprising and, you know, our producer, Adam, really wanted me to bring this up.
And that is the model announcements they made at this were surprisingly small to me.
You know, they did announce Gemini 3.5 Flash, which they are claiming is just like one of the fastest models out on the market.
And I generally believe them because, you know, you can see it in action, like in Google search and see how quickly that model responds to you.
But what was really missing was like something to compete with, you know, the other big players in this space, like Anthropic and OpenAI.
They didn't really announce any new frontier, bigger thinking models, which is kind of interesting because, you know, the last time they made big announcements around models, it felt like a flurry of everyone trying to like, you know, to hurdle back over Google.
So it's almost like they're taking a different approach this time.
And, you know, maybe this is a sign that efficiency is now the thing that all of these companies are going to start focusing on.
But yeah, what do you think about all these other announcements, Andrew?
I agree with you about, you know, the...
A big splash of a huge frontier model being missing kind of definitely alters the atmosphere of the event.
But it allowed Google to give more attention and focus to these other star players of their stack that really helped them stand apart.
And they don't need, this is the key thing.
They don't need the stage of Google I.O.
to drop their next flagship model.
They could drop that on a random Monday morning and we're all still going to be falling over.
over ourselves to go test all of our workflows with it, to go see what new tools extend it first.
All of our inboxes are going to blow up with those, you know, 50 plus emails of every tool that we use, letting us know that the latest model is available now for them to try out.
So, you know, we will already be sat.
And so we don't need Google I.O.
to tell us that.
And that was a smart choice on their end to then let's direct this attention to the stuff that maybe doesn't normally get as much eyes as it should.
All right, Angel, let's talk about this research team that could use the help of our audience.
What do we have here?
Yes.
So on our show, we've talked about the phenomenon of dark flow with vibe coding.
We talked a few months ago about it was a meter study about the effects of vibe coding, the flow state that results and how that's comparable to other states like gambling and being at a machine and losing a sense of time and having a false installation of confidence of your influence on what's coming out of it.
So this was an initial.
survey that kind of helped understand and identify these similar patterns.
Well, Dev Interrupted is actually partnering with some researchers that are doing a more in-depth study on the effects of vibe coding and flow state on developers and engineers.
And they're looking for respondents to their survey.
It just takes about 10 minutes or so.
And if you're somebody who's using any of these tools, you're likely qualified just by listening to us.
We'll share a link below.
And if you can go and participate, we'd really appreciate it because this is our ability to not only report on the stories that define our practice, but also participate in the research to further understand them.
And we're going to bring all of y'all along for the ride as well, because we're going to be part of that research process.
So check out the links and definitely reach out.
Yeah, I love to be able to help our research teams like this.
And, you know, I think it's important to put things in context.
And right now we're sort of in a moment in history where a lot of how we work is being transformed by this new technology.
And one of the important things that we can do along the way is sort of document the state of things so that we have a clear understanding of how.
society and ourselves are changing, you know, because I certainly feel different when I'm under the influence of agentic coding.
And it would be more interesting to know more about what that mental state means for me and for society as a whole.
So yeah, definitely check it out if you have a few minutes.
Totally.
All right, Andrew, let's talk about distilling yourself.
What do we have here?
All right.
This is a great article from James Stanier, who we've covered on the pod before.
James reporting from an engineering leadership vantage about helping equipping those leaders with like, how do they tackle working with these tools at scale?
Because a really interesting thing happens when you use AI tools in a more managerial position is if you don't take the time to instill the values of your company and what you should be aligning to, if you don't take the time to boil down the distinct perspective and the things that you care about, then ultimately what comes out is generic.
We all know this, and this is a problem for ICs trying to get their just daily work done, but this becomes an even larger amplified.
problem for managers if they are unable to provide that specialization and taste and judgment to their AI workflows.
That way, when they influence their downstream projects and direct reports that not only aligned, but they're only saying what needs to be said.
This article tackles the idea of distilling yourself, the idea of doing an interview style rounds with the LLM that you work with, giving it examples of past work that you do on a regular basis.
And instead of assuming that you know how you do your work done, put it in the position of interrogating you and digging into your practices and developing almost like a profile on what makes you the manager you are and what you care about.
This is a practice, obviously, everyone should.
engaging with.
It's how anybody actually manages to leverage and distill their taste into the workflows that they do.
But this is just clearly naming it and giving you some practices.
So Ben, what stood out to you about this one?
You know, I think it's a really great way to think about if you're an engineering leader who's, you know, wanting to leverage AI to be a better communicator.
You know, a lot of the focus of this article is on how do you...
leverage the system for writing, you know, whether that be a blogger or someone who publishes to a newsletter, or you just need to be better at communicating requirements and important things to your team.
One of the really awesome things about AI is that it is opening up opportunities for all of us to broaden the reach of our capabilities.
So, you know, for example, I'm doing things as a part of my daily work that I wouldn't have considered in the pre-AI era because I didn't have the skills, nor did I have the time to develop them, even if I could quickly understand whether or not what I was doing was good or bad.
You know, one example is social media.
Like I just I hate writing social media stuff.
But AI, if you can frame your thoughts to it in a really great way, it can help you sort of format it for the format that you're trying to publish it to.
But, you know, every sort of communication challenge can be solved in very similar ways.
You know, and it admits like all of this agentic coding fervor.
I'm actually meeting more and more engineers who are actually deciding to develop their writing and coaching skills specifically with AI.
And I think it's actually a really can be a really great practice if you do it right.
And, you know, a big thing that James mentions in this article is how when you write things down, it actually forces you to apply really deep and critical thought to your work.
In a way that nothing else really replicates.
And this is really what has always drawn me to the craft because it really does help me understand my perspective on the world.
And a big part of that is that written word needs to be defensible because it lives in this semi-permanent state going forward.
So if you're feeding it into an AI system, it's going to keep reverberating back into that AI system over and over again.
And, you know, I think writing is a skill that anyone who wants to be a better leader should work on developing, even if you're in software engineering.
And using AI to distill yourself into an instruction set for AI, you know, that's how you get AI to like work the way that you want it to.
But then more importantly, you have to use it, you know, all the time and build that communication muscle.
So yeah, go check out this article from a friend of a show we featured in, I know, in the past.
And if you're an engineering leader out there who's looking for to learn just how to communicate better with your team, I think this is a really great place to start.
I agree.
All right.
Let's talk about this open source alternative to Obsidian.
So we saw this new open source project called files.md that came across our desk.
And it's a minimalist note-taking app that runs entirely in the browser and stores everything as local MD files.
And I'm like, wonderful.
I love that.
I am totally local now with MD files.
So that describes me to the T.
But it actually focuses a lot on simplicity, or it focuses exclusively on simplicity, to the point where the creator of this argues that this second brain complexity that we've all started to adopt in the AI era is...
It's just a procrastination tool that defers actual thinking rather than, you know, instead you should advocate for your first brain to actively process and connect ideas rather than just collecting them.
So, you know, this little app incorporates a bunch of really cool features that just make it easy to capture all of your thoughts in local Markdown files.
And definitely worth checking out if you're trying to get into this.
this habit of building context to your daily work.
So first of all, I love the minimalist markdown tools that are all out there right now and the local first design.
That is my life right now and I think it's cool.
But I actually do kind of love the complexity of Obsidian, particularly as I've gotten pretty far along with my developing NodeGraph.
I don't know, maybe it's just like I've invested too much now and I won't give up, but I'm starting to see the benefits of it.
Even if I can understand, like, you know, there's some, I can understand some of the concerns of this, particularly around, like, making sure you actually process ideas rather than just capture them.
You know, I do think that is a very important practice.
But I don't know, Andrew, are you using files.md?
So there's a lot in this to unpack that, you know.
I want to start by saying that this is really the vehicle by which you apply what we talked about last article.
So in the last article, you're thinking of like distilling yourself, then Obsidian or Files.md becomes the place where that distillation lives.
And there's a few important things about this.
that are worth calling out.
These are the durable primitives you should be betting on.
One, local first, being able to own and have all your own files locally, that is a baseline need.
And if you need them accessible in more than one place, this is when you introduce source control.
You don't want to solve it with an abstraction.
You want to use a durable primitive that's already there.
So you would use source control.
And as well as how can you make that tech, those text files as future proof as possible.
Well, you have two options.
You can make them a .txt or you make them a .md.
Why?
Because those are files that are always going to be openable, that you can always interact with, that are going to be the smallest versions of a text file that you can use as an I.O.
on any other kind of workflow that you want.
So you don't want to have it locked away in some abstraction, like a notion table or like locked away in a very special.
specialized like note-taking app where sure you could go and hit that export button but will you the thing is is that you won't and that's where I loved what this author doubled down on the idea of well don't go building your second brain and then ignore your first brain there's a lot of activities that happen once you pick up this kind of practice that kind of just turns into arranging the deck chairs on the titanic of your own mental catastrophe because you end up spending all of this time manicuring and building this system that contain your knowledge and none of the time that you should be thinking about the knowledge that's sitting inside of it.
So don't build silos that you won't use.
And ultimately, you should bet on durable primitives.
So to answer your question, Ben, am I using files.md?
I'm not just because I'm using Obsidian.
But the great part is, is that neither you or I are locked into Obsidian.
So if we wanted to go try files at MD, we just point this thing at that same durable archive of markdown files that we keep.
And even going back to what you said of, oh, you love Obsidian because it has a graph.
All those graphs are connections of links from those markdown files to other files.
The abstractions by which we understand the relationships of those, those are solvable.
You could use that in a folder and ask Claude to give you that same graph on demand whenever you want to see it.
And you have to keep that in mind when working with these tools.
Don't bet on the shiny thing that captures your attention, but bet on the primitives that won't lock you in and will grow with you.
Yeah, man, kind of bringing this full circle to my comments about, you know, needing that collaborative agentic document editor.
You know, you can use both of these obsidian and files at MD side by side.
You know, sometimes you want to be more in a focused mode and like working on cultivating that first brain.
And other times you need to be more on a network mode where you're collecting thoughts and like bringing them together and to analyze it collectively.
It's a fun time because I feel like we're just finding new ways to interact with all of these artifacts that we're generating through our daily life that is changing a lot.
And it's pretty exciting.
Yeah, it's definitely an interesting time to be experimenting with these kinds of tools.
If you haven't picked up something and started experimenting with your version of a second brain, this is your weekly reminder to do it.
I think Ben and I talk about this phenomenon almost on a weekly basis now.
And it's really going to become a differentiator for leaders, especially those within very specialized fields, because the ones that can crystallize that specialization into these repeatable processes are just going to work.
work faster than those who don't so andrew what are your agents up to right now oh that's a good question what are they up to i've left them alone for a few minutes that's actually kind of scary for you to say what they might be working on okay so oh currently right now my agents are going through the skills olympics on my machine Because I've collected so many skills over the last few months.
I actually created this arena where I've dumped in a bunch of skills files that I found from different sources, different ones I've been using.
And I've scoped out these different projects with different arrangements of them.
And I've given them some tasks.
So I have some agents that have everything, some agents that have very little.
And what I'm trying to understand now is how much of my process is baggage.
And for the things that are somewhat useful.
but maybe don't belong on a global level, where should they live?
I think skills management is a really interesting emerging part of the agentic space because I don't know about anyone else, but I get really, really excited when I have an opportunity to boil something down into a skill that I can carry with me forever.
It's like you just put the little feather in your cap and you know you're going to be able to use it again whenever you want with whatever flavor.
It's really quite fun to go and try to capture them.
But if you're like me, you probably have a huge pile of them.
So what do you do?
If you are experimenting with different ways of managing your skills, I'd love to hear from you.
But what about you, Ben?
Yeah, that's another great primitive that has to be solved.
And I don't know, Git kind of works for that, but I think we need something better.
Yeah, my agents have kind of been all over the place.
I've been going way horizontal than I typically do.
You know, it's been really fun to just like, you know, use these different tools to like adopt new practices that I have not, like, I understand what a good outcome looks like, but I don't necessarily know what it takes to build it.
And, you know, as long as you can keep the problem constrained enough.
I can apply my second brain to this and generally get the outcomes that I want, which has been pretty, it's just, it's wild, you know, to see, you know, the different, the breadth of capabilities that AI can sometimes grant you.
So yeah, I've not really been operating in an agentic mode per se, more so in a, I guess I'm really more like context engineering right now.
Like I'm just trying to collect all of the...
the important context that I need so that I can start to, to give agents enough information to go start making those decisions.
So yeah, I feel like there's an ebb and a flow to it.
You know, sometimes you're just ordering agents, what to tell them what to do.
And other times you got to plan for, there's an incredible amount of stillness and knowing exactly what to do.
And sometimes you just have to get wait until you're perfectly aligned and then you can strike.
Yeah, exactly.
All right.
All right.
Well, thanks everyone for joining us today.
That is the Friday Deploy presented by Linear B.
Give us a like wherever you're listening to us or a comment, drop a comment, a rating, whatever you can do to help us.
You know, all this stuff helps us grow the show.
So we appreciate the help.
Thanks for joining us and we'll see you next week.
See you next time.
Are you looking for a trusted way to evaluate engineering productivity platforms in the AI era?
Gartner just released the first ever Magic Quadrant for developer productivity insight platforms, and Linear B was named a leader.
As AI changes how software gets built, engineering leaders need better visibility into productivity, bottlenecks, and AI ROI.
Download your complimentary copy of the Magic Quadrant to see why this category matters now, how the market is evolving, and why Linear B is recognized for its vision, execution, and workflow automation.
Check the show notes for the link.