# MCP Certification and AI Velocity Paradox

**Podcast:** Dev Interrupted
**Published:** 2026-07-17

## Transcript

I was at this event this week and I was talking to, you know, a whole bunch of geeks about AI and all this stuff.
But yeah, it's not surprising at all to me how many people out there are launching their own game company right now with AI.
I don't know.
It's not very surprising.
Yeah.
Yeah.
Not at all.
I mean, I launched one over the weekend with my five-year-old.
I met another fellow dad, was building games with his kids.
I loved it because he had like educational versions of like missile commands, for example.
And then he's like already sent me the link where I can just go and like sign up and have my kids start playing this.
So it is truly an exciting time.
No, I mean, software now is just becoming a communication method, an easier way to like have fun with folks and the collaborate, explore ideas, especially.
like with kids as well, with like just with their imagination.
But even just using AI to augment the things that they love or are interested in is an amazing opportunity, I think, for folks, especially those that are maybe like educating or mentoring younger ones.
Like a recent guest we had on the show in the last year.
He was telling me all about how he built like a hockey field thing in his backyard with like a drone that had like all this like stuff on it.
Basically, that became like an overlay of all of his kids like hockey games they play in his backyard.
And so they have like a scoreboard and the neighbors could watch.
And so it became like a whole like neighborhood bonding thing around all the local folks who love to play hockey.
And it's just all augmented by technology he was building for fun while being a curious software leader.
So I think just, you know, finding.
unique ways to bridge it back into our daily lives is something we're going to be stumbling into a lot more of.
Yeah.
And I'm curious, as we've all been out there building our own game companies and drone services and sports arenas, have you found is record in your code yet?
Has it shown up in your code base that you've noticed?
Oh, yes.
So this is actually a classic conundrum I've seen people talking about on social media about an is record checker inside of a code actually specifically produced by AI Codex.
And since I'm kind of more of a cloud code guy, I actually haven't stumbled across is record in the wild.
But I'm sure maybe some of our listeners have.
But effectively, all this is is a type guard checking whether an object is actually an object.
It's kind of just like a...
zero or a one introspecting on itself.
And it's not a very useful piece of code, especially when you cruft it all up as something as abstract as is record.
And when you scroll past it in your type definitions or in your guards, you might even completely miss it.
But I promise that this definition might be hiding in your code somewhere.
There was a really fun article that came out this week from TLDraw.
It's an open source Canvas app.
And they talked about how they're...
The isRecord code was back in their code base in 2022 as part of their data store package.
And yeah, maybe it wasn't a very useful piece of code then, but it does seem like at least that chat GPT and OpenAI world to have a fixation on this thing that it found in their repo because it's since shown up in a bunch of places.
And what I think is really fascinating here is...
It raises the, it brings rather to the foreground a question and observation, a thought we've had in the back of our minds about, you know, open source software, even closed source software, proprietary things are all underneath the model and power its capabilities and are part of its training data.
So what does it mean when these things pop up and what can we learn from it?
And what other things like is record might be popping up in places like our code?
Yeah, yeah, this I found this story to be very charming, particularly as somebody who sees like echoes of myself in other people's AI usage, you know, things that conventions that I come up with that get copied into someone else's AI workflow and then just continues to get echoed everywhere.
But you know, it's kind of funny, because like, ever since like the dawn of the internet, we've all had to be like conscious of things that we put on the internet, because it could become like a permanent fixture, basically forever.
But now, like anything you publish can become a permanent echo that just like resonates on onward and onward.
Yeah, the consequences of that are even more extreme.
It makes me think of like real physical problems like we've polluted parts of like our upper atmosphere that make it hard for satellites and spacecraft to move through it.
And we could eventually be in this position where there's tons of garbage.
floating around an orbit of Earth that make it impossible to have safe space transit one day.
And it's like the idea of all of this stuff becoming a permanent consequence that we have to deal with later becomes even more staggering when you think about training data.
And it's not just an embarrassing photo that you put on MySpace a bunch of years ago, but it's actually maybe everything you've ever written.
Yeah.
Well, anyways, let's get to the point.
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 this week, we're covering MCP getting its first certification, getting started with agentic loops, taming your agent's toolbox, rewriting products from the ground up for AI, and the strongest data yet that we've seen on the velocity paradox of AI.
So let's just start right at the top with MCP getting its first official certification.
What do we have here, Andrew?
Yes, so Linux Foundation's Agentic AI Foundation, they're launching the Model Context Protocol Associate.
This is the MCPA.
It's the first of its kind official certification for MCP, the protocol that's rapidly become the standard for connecting AI applications to external tools and data.
Beloved child of the hype cycle who's been killed and brought to life about five times at this point, and MCP is finally getting its first official training course.
So for engineering leaders, this opens up a new opportunity to screen and understand the skills that go into managing and deploying these external tools at scale for agents, especially as enterprises figure out the best ways to use and deploy MCP, which is arriving at its 1.0 spec finalization later this month on July 28th.
It's a forward compatible, backwards compatible spec change that doesn't...
really break MCP servers for pretty much any use case except some very obscure ones.
So it's an opportunity to get rid of lots of cruft in the spec and finally train folks on the best way to be managing these tool providers moving forward for agents.
The exam was developed with input from a whole bunch of cross-industry collaborators.
This speaks, I think, to the universal importance of open specs and protocols and standards like MCP.
And among them is Anthropic and Google, AWS, Microsoft, GitHub, Hugging Face, many names that we've had on this show and that you recognize that power all of the AI technology we use every day.
So there's a lot of cross-industry credibility here.
So definitely be sure to check it out.
It's not live yet, but it will be.
be available very shortly and they're making registrations available soon.
Yeah, you know, there's been a lot of back and forth on MCP over the last two years, I would say, about whether or not it is a permanent fixture or it's just some sort of temporary stopgap as we get to the next thing.
It really does seem like it's starting to earn its right to exist, you know, and it is like becoming a very valuable tool in the agentic tech stack.
And I've always kind of looked at the Linux Foundation as like, you know, when they view something as being certification worthy, so to speak, you know, I think that's a good signal that this technology is really starting to get to a mature state and, you know, enterprise ready and ready to adopt.
So we've had MCP within the Linear B platform and we sort of started doing it back when it was pretty experimental, I would say.
And, you know, it was clear from a...
very early on that it was a bit of a wild west, you know, at the start.
And, you know, I'm definitely looking forward to having better standardization and capabilities around MCP.
So this is great.
Absolutely.
All right, Andrew, let's talk about getting started with loops.
This is a concept that we talk about a lot, but I think we, you know, I wanted to share this article because I do think it's just a really great breakdown of like, where our understanding of this.
concept is and how, you know, engineering leaders out there everywhere can take advantage of it.
So what the hell is a loop anyway, Andrew?
Oh, gosh.
Well, a loop is what's making all of this agentic engineering exponential growth and compounding stuff that we're seeing on charts and, you know, we're being responsible for all of the yellow and orange status pages of all of our beloved services we use every day.
Loops are ultimately the closed system in which you provide feedback to a system that otherwise self-propels itself to get your work done.
And this is a pattern that is simple on its face.
It's effectively a loop.
It's a for loop.
It could be any number of gates and guards along the step of finishing that loop.
But ultimately, the goal is for it to do it again and again and again and for you not have to be there.
Ultimately, this is an engineering goal that many people are working towards.
with their agents, with their orchestrators, to get them into a position of proactively continuing and finishing their work, doing long-horizon running tasks that maybe even take much longer than typical days to finish.
We're talking about agents that can keep a durable state in memory and can remember and improve between those iterations of the loop.
This is a new level of engineering that folks are still getting kind of comfortable with, I think, at the top of the year.
we were all getting comfortable with the idea of having the orchestrator, right?
The top level agent that understands how to command and summon the sub-agents to get all of the work done.
That aggressively protects its own context to get the job done.
And now what we're seeing is the orchestrator itself, you now probably have a whole bunch of orchestrators that all work on...
different specialized things.
Really what you need to do now is connect them together, create a feedback loop in which they can support and help each other, unstuck each other, but otherwise also include you in the critical steps of what means the work gets done.
And so for many folks, this means like at this point, if you have a bunch of agentic capabilities you've built up and you're able to kind of effectively press the gas pedal on your client or harness of choice and get the work done, this is your opportunity to reflect.
When I press that gas pedal and I tell my agents to go forward, what am I doing?
Am I presenting a problem?
Am I providing feedback?
And then finding ways to make that process happen either autonomously with guards in the background or surfaced to you proactively.
That way you can push the loop forward instead of having to drag it to get it started again.
And the difference in that momentum is profound once you find the gaps in doing that.
wanted to call out that this is like a really insightful way to dive into loop engineering as it is yeah the world really just is loops all the way down at this point or is it sub-agents It really feels like a level of zooming out.
Like when we were first using agents and you're getting autocorrect and then you're getting it to write its code by itself and then you're not even looking at it.
And then you have multiple of them.
You're the orchestrator.
And then you make the orchestrator.
And then you're just watching the orchestrator.
Now we're zooming out even more.
Your orchestrator has orchestrator neighbors.
And you have a whole system that's getting really complex to manage.
So what do you need to do?
You need to do what a CEO would do.
Hire the manager of those people and manage that person instead.
This is all about empowering our own org charts.
And the loops are really just the representation of that.
Yeah.
So whether your loop is an orchestrator, a sub-agent, or a sub-agent of a sub-agent, I think one thing we're really learning from all of this is that context is everything.
We've been talking about context engineering more and more over the last year or so.
And I think what we're really learning is that at every layer of this loop, you have to have all of the context that is necessary to be successful injected into it.
And that's kind of the brilliance of breaking it down into this loop flow is that you're chunking off work in a size that you could identify what context needs to be pulled in to either make better decisions or know when you need to bring the human in as you were calling out.
That's something we've really been exploring at Linear B.
As we're seeing engineering teams everywhere, just adopting AI and then starting to go full agentic in many situations.
There's just so much information that you can feed into those agentic loops that help them make better decisions and move faster and alert their humans when something is at risk of going off the rails.
So yeah, this article is a really great way if you've just woken up.
from a coma or you got out from under your rock or something and you've not heard of loops, or maybe you just want a refresher on like, on how we've gotten to where we are with this concept and where it is today.
This is a really great article that just breaks everything down.
All right, let's move on to a topic that we're relatively familiar with right now, but it's cool to see another company out there doing this.
So what's going on with CircleCI and why are they rebuilding their CLI tool from scratch?
Okay, this one's going to be a tongue twister, but CircleCI is rebuilding their CLI.
And this is an innovation that is targeted towards their new primary users that's emerging, which is obviously agents.
I can't think of a better consumer of agentic tools or a better user of agentic tools than CICD because it's such a virtuous part of getting work done and contributing to a larger code base.
And it's a critical part of, especially when you're working with agents and making sure that they're not clobbering your code.
workers and that the work that you and them are producing is actually up to standards.
So this is the idea of redesigning the CLI.
It's built in Go with a whole principle being human first, but with a composability opt-in.
That means that it renders cleanly in the terminal, but it outputs or pipes plain data when asked to.
And if you've been using CLI tools, this is probably familiar to you as like a dash dash JSON kind of argument.
The idea that anything you could see as a human and from an agent in the terminal can be turned into delicious structured data for an agent to very quickly and rapidly use.
This is effectively turning the CLI into a hybridized MCP server because the agent can use it just like the user can.
And while this maybe sidesteps some of the parts of MCP that help with distribution or otherwise authorization, this does bring into the developer seat and into the agentic development workflow the ability to run CICD locally and get results, allowing for more rapid iteration and conformant to the standards that make for good software for that particular work.
It also just has a bunch of tooling built in running pipelines and jobs at scale and debugging failures right there in the terminal so that your agent could do it for you.
It unlocks new opportunities for agents to iterate and improve on the CI CD in your world of CircleCI and honestly be a stronger collaborator with you and your team.
So a really great direction to see this tool go.
I think this is going to be a trend you see for a lot of your favorite tools, especially ones that sit right in the developer's IDE and is something near and dear to every PR, is that they're going to get moved further and further left in the development process.
And right now, there's nothing further left than the agent.
So CircleCI is running right there to be with them.
Yeah.
And, you know, normally we don't cover like product updates like this, but I did think it was pretty cool.
It was a story that was pretty relevant to us and just interesting from the perspective of, you know, an established company that has to adopt or adapt to this new era.
You know, it's very similar to what a lot of the changes that we've been going through at Linear B as well.
You know, we've we started as sort of like this visibility layer for engineering leaders, you know, to get all their productivity data, which we still have all of that.
But, you know, it really all started to change a little bit.
It really started to change when we introduced MCP to our product, you know, and I mentioned that a little bit earlier, but this changed how people interact with the value we provide on a very fundamental level, because instead of having to learn how to use a platform and go out and, you know, either find the things that answer your questions or build them yourself or, you know, do an analysis to try to understand your data.
Suddenly you're just asking, you know, your AI tool of choice to connect to our MCP server.
And to build whatever it is that you need, whether it's a dashboard to report to your executives or if it's just some you want to understand what's slowing your team down right now.
And it really did start to fundamentally change how people interacted with our platform.
And that's only continued to increase.
You know, we've been introducing more features around having natural language interfaces into our platform and letting those interfaces build the platform itself, too.
And all of this is just opening up new possibilities and what you're able to do with the platform.
You can combine it in ways, you know, all of this data that we have in ways that just wasn't possible before.
But then it's also just so much simpler now.
And we've had to like really rethink, like, how do we present our product to an audience that, you know, today it's a human that's asking these questions.
And getting data out of it.
But we very soon anticipate that it will be those agents operating in those loops that will be constantly looking for data sources like what we have or like what CircleCI has.
So, you know, I think if you haven't felt this type of transformation happening yet to your company, it's probably only a matter of time.
But and if you have, then, you know, just know that everyone's sort of going through this right now.
So, you know, we can learn a lot from each other.
Yeah, a really cool thing from CircleCI, and I'm looking forward to watching for more news from them.
Yeah.
So our last item here, Ben, is some research.
Do you want to tell us about this interesting research article that came across our desk this week?
Yeah, of course.
File this whole way under research that just, again, affirms linear B and all the things that we care about here.
But this is a study that showed that AI writes faster than humans can review.
And it's a longitudinal study of a 2x AI mandate at a large enterprise.
So this study analyzed over 800 developers across about 200,000 pull requests and found that a company-wide mandate to double the merged PRs per engineer actually resulted in double the PRs coming out.
Making this actually a pretty large example of validated.
AI coding gains that are happening out there.
So, and they noted that, you know, the gains were broadly shared across seniority levels, which was a pretty interesting fact, but the gains were also concentrated largely in newer code, which really isn't all that unsurprising.
So if you're in like a legacy code base, you may not see like the same thing.
Now, you know, there's all debates about whether or not doubling output is actually the thing that you should.
you should set goals against and, you know, it's debatable for sure.
But, you know, we here at Linear B have some engineering benchmarks or AI benchmarks that we're going to be releasing here very soon that, you know, have really confirmed a lot of what we're seeing in this research as well.
So, you know, across the board, it does seem like AI is now starting to result in significant increases to output.
So, Andrew, what do you think about this research?
Yeah, something that really stood out to me that was interesting is that the per reviewer workload, it...
doubled as a result of, you know, that doubling of the actual output of PRs.
And so what happened was automated review then overtook human review within the organization.
So you're talking about an organization with 800 engineers that have just now doubled the amount of PRs they make.
And then now on top of it, the automated review that's pushing all of those through, you know, it doesn't necessarily involve them at least too closely in the loop.
It's largely doing the custodial checks that it needs and it's saving them a lot of time.
time as well.
But what happened was that the automated review in some capacities, at least at first, it overtook the volume of what they could handle.
There were definitely the SDLC struggles under that volume of AI-assisted.
code where like a lot more comes into being that can be reviewed.
And so things like AI code review are really the solution for unblocking that.
And so I think what this article or rather what this research really kind of synthesizes is that, you know, shipping more code.
achievable, easy, actually.
This is a large organization that doubled how much code they were shipping.
But like the review and quality infrastructure that is needed to scale with it introduces a lot of risks and bottlenecks that can offset the throughput gains unless you have a clear strategy and understanding of what code you're shipping and what is the ROI on the tools that you're using.
Yeah.
And if you've listened to Dev Interrupted long enough, I probably sound like a broken record on this topic, but we've covered endlessly about how code reviews are the most common bottleneck within the typical engineering organization.
And AI is amplifying everything.
So if you have a code review that's already overwhelmed by too much code, then AI is just going to make that worse.
And that's really, linear B, we spent a lot of time focusing on that just because it is such a common.
phenomenon.
So, you know, it's why we've built our GitStream workflow automation for the code review process.
And it's why we built an AI code review tool because, you know, to give another preview of our upcoming updates to our benchmarks, you know, code review, an AI code review is one of the easiest ways to increase the amount of code that you get out into production today.
So yeah, it's great research that just affirms, you know, everything that we believe in.
And I always love to read that.
Yeah.
Well, thank you for sticking around to the end of this episode, which is brought to you by Linear B, the engineering productivity platform.
We're seeing a widening gap between engineering teams that have turned AI adoption into delivery and those that haven't.
And it's real and it's measurable.
And in fact, we have a workshop coming up on July 30th, where we are going to cover this exact topic, how you have these cohorts of engineers who are operating in a highly agentic manner that are...
moving ahead of the pack in terms of code output to production.
So in this workshop, you'll see why AI usage correlates with a doubling of PR merge rate and why more AI code actually may not mean more shipped code.
So thanks again for joining us for this episode.
Be sure to give us a like wherever you're listening to us.
Rate this podcast or the YouTube show if you're listening to us over there.
And make sure you register and join us on July 30th for this workshop.
So thanks again, and we'll see you next week.
See you next time.