Agentic Engineering: Why Trust Became the New Bottleneck

WEBVTT
00:00:02.399 --> 00:00:14.080
To Digital Pulse, I'm Rafa, producer and host of this series where we explore how emerging technologies like digital assets, blockchain, and AI drive real innovation in business.
00:00:14.320 --> 00:00:18.800
Each episode is crafted using a fully AI-enabled workflow.
00:00:19.039 --> 00:00:23.600
I'll now hand it over to my AI co-hosts to take you through today's episode.
00:00:24.000 --> 00:00:24.719
Thanks, Rafa.
00:00:25.039 --> 00:00:27.359
Welcome everyone to another episode of Digital Pulse.
00:00:27.760 --> 00:00:28.320
Glad to be here.
00:00:28.399 --> 00:00:30.399
This is uh such a critical topic today.
00:00:30.719 --> 00:00:31.359
It really is.
00:00:31.839 --> 00:00:35.439
So I want you to imagine getting a knock on the door from a federal auditor.
00:00:35.840 --> 00:00:38.640
Oh, the absolute nightmare scenario right off the bat.
00:00:38.960 --> 00:00:39.119
Right.
00:00:39.200 --> 00:00:40.399
Let's just set the stakes.
00:00:40.560 --> 00:00:48.640
They walk into your office, open up a laptop, and they point to a critical risk decision module in your core banking system.
00:00:48.960 --> 00:00:49.119
Okay.
00:00:49.679 --> 00:00:51.679
And they ask a very simple question.
00:00:52.000 --> 00:00:56.320
Who wrote this logic and why was this specific threshold chosen?
00:00:56.399 --> 00:00:58.719
Aaron Powell A completely standard compliance question.
00:00:58.799 --> 00:00:58.960
Trevor Burrus, Jr.
00:00:59.039 --> 00:00:59.280
Exactly.
00:00:59.439 --> 00:01:04.079
So you look at the git blame logs and the answer is essentially, well, nobody.
00:01:04.400 --> 00:01:04.560
Right.
00:01:04.879 --> 00:01:08.079
An AI generated it during a quick chat session on a Tuesday.
00:01:08.239 --> 00:01:12.159
Your lead engineer, you know, glanced at it, saw that it compiled, and just hit approve.
00:01:12.239 --> 00:01:12.400
Trevor Burrus, Jr.
00:01:12.480 --> 00:01:13.840
Which happens every single day right now.
00:01:13.920 --> 00:01:14.879
Aaron Powell It does.
00:01:15.200 --> 00:01:18.640
But in a regulated industry, that answer isn't just embarrassing.
00:01:18.719 --> 00:01:19.519
It is illegal.
00:01:19.599 --> 00:01:20.319
Trevor Burrus Completely.
00:01:20.719 --> 00:01:27.680
It's the kind of answer that results in massive fines, executive panic, and honestly, the potential loss of your operational license.
00:01:27.760 --> 00:01:29.760
Aaron Powell Yeah, that is not an exaggeration.
00:01:29.920 --> 00:01:31.920
That is the reality we are stepping into.
00:01:32.000 --> 00:01:42.239
Aaron Powell So today we are talking directly to you, the enterprise tech leaders, the chief technology officers, the chief information officers, and you know, the senior buyers of software services.
00:01:42.319 --> 00:01:48.319
Aaron Powell Because you are all sitting at the epicenter of this massive, extremely dangerous shift right now.
00:01:48.719 --> 00:01:49.439
You really are.
00:01:49.680 --> 00:01:54.719
You're being bombarded every single day by vendor claims about artificial intelligence.
00:01:55.120 --> 00:01:56.719
Oh, the sales decks are just relentless.
00:01:57.040 --> 00:02:04.079
Every single one promises that AI will revolutionize your delivery, cut your costs to zero, double your deployment speed.
00:02:04.480 --> 00:02:06.959
But here is the thing: you are the one carrying the liability.
00:02:07.359 --> 00:02:07.760
Exactly.
00:02:08.000 --> 00:02:14.400
You are the one who has to figure out how rigorously to actually adopt this technology within your mission critical teams.
00:02:14.719 --> 00:02:17.360
And it is an incredibly precarious position to be in.
00:02:17.520 --> 00:02:21.280
I mean, the noise in the enterprise software space right now has reached a fever pitch.
00:02:21.360 --> 00:02:21.520
Trevor Burrus, Jr.
00:02:21.599 --> 00:02:23.039
The signal is almost completely lost.
00:02:23.360 --> 00:02:23.919
It really is.
00:02:24.080 --> 00:02:33.520
There is a very real quantifiable danger of adopting AI in a way that just creates massive long-term liability for your organization.
00:02:33.599 --> 00:02:40.319
Aaron Powell Like we're seeing companies buy hundreds of enterprise licenses for conversational coding assistance, right?
00:02:40.400 --> 00:02:46.080
Aaron Powell Yeah, believing they are modernizing when in reality they are just industrializing their technical debt.
00:02:46.159 --> 00:02:47.919
Aaron Powell Industrializing technical debt.
00:02:48.080 --> 00:02:49.120
That is exactly it.
00:02:49.280 --> 00:02:52.159
So we are moving past the introductory concepts today.
00:02:52.240 --> 00:02:57.520
Aaron Powell We are looking at the structural operational reality of building software in this new era.
00:02:57.599 --> 00:03:02.000
Aaron Powell Which brings us to the core concept we need to unpack today agentic engineering.
00:03:02.319 --> 00:03:02.560
Right.
00:03:02.719 --> 00:03:07.520
And we need to define this carefully because the market is throwing that term around very loosely right now.
00:03:07.919 --> 00:03:11.759
So for those of you evaluating vendor claims, let's get perfectly aligned.
00:03:12.000 --> 00:03:16.159
When we say agentic engineering, we're talking about specialized AI agents.
00:03:16.240 --> 00:03:16.800
Aaron Powell Exactly.
00:03:16.879 --> 00:03:18.400
And the key there is specialized.
00:03:18.800 --> 00:03:19.039
Right.
00:03:19.199 --> 00:03:25.280
You have one specific AI agent dedicated to every single role in the software development lifecycle.
00:03:25.599 --> 00:03:28.159
One agent acts as the analyst, another as the architect.
00:03:28.400 --> 00:03:30.479
Another as the coder, and another as the tester.
00:03:30.800 --> 00:03:37.199
And crucially, and this is the most important part, these agents are working strictly under human supervision to build enterprise solutions.
00:03:37.520 --> 00:03:37.840
Yes.
00:03:38.080 --> 00:03:43.039
This is not about a developer just uh using a smart auto-complete tool in their IDE.
00:03:43.360 --> 00:03:44.159
No, not at all.
00:03:44.400 --> 00:03:51.520
Agenic engineering is a complete foundational re-engineering of the organizational chart and the delivery pipeline.
00:03:51.599 --> 00:03:54.240
Aaron Powell And that structural distinction is everything.
00:03:54.400 --> 00:03:58.960
I mean, the core premise of our discussion today flies directly in the face of the current hype cycle.
00:03:59.120 --> 00:04:04.639
Aaron Powell Which is that casual use of AI in software development actually accelerates your delivery.
00:04:04.960 --> 00:04:05.120
Right.
00:04:05.360 --> 00:04:05.759
It doesn't.
00:04:06.000 --> 00:04:08.879
Casual AI doesn't accelerate delivery, it just produces faster failure.
00:04:08.960 --> 00:04:09.919
Aaron Powell Faster failure.
00:04:10.159 --> 00:04:10.400
Wow.
00:04:10.800 --> 00:04:13.680
Because the bottleneck for enterprise software has fundamentally moved.
00:04:13.759 --> 00:04:14.479
Aaron Powell It has.
00:04:14.639 --> 00:04:19.839
For four decades, the bottleneck was execution, the physical manual act of writing the code.
00:04:20.000 --> 00:04:21.439
Aaron Powell Sitting there typing syntax.
00:04:21.759 --> 00:04:22.079
Exactly.
00:04:22.319 --> 00:04:23.439
But that is no longer true.
00:04:23.600 --> 00:04:26.720
The bottleneck has moved from execution to trust.
00:04:26.800 --> 00:04:30.079
Aaron Ross Powell So today we're going to map out that transition.
00:04:30.319 --> 00:04:35.839
Moving from viewing AI as just a coding accelerator to implementing AI as a structured operating model.
00:04:35.920 --> 00:04:42.560
Aaron Powell And I want to dig into this new bottleneck immediately because it really challenges the primary illusion holding the industry captive right now.
00:04:42.720 --> 00:04:48.480
Aaron Ross Powell You mean the idea that faster coding automatically equals better enterprise delivery.
00:04:48.560 --> 00:04:49.120
Aaron Powell Yeah.
00:04:49.439 --> 00:04:51.920
The entire market seems hypnotized by this.
00:04:52.160 --> 00:04:57.360
Vendors run these demos showing an app being built in three minutes, and you know, executives are applauding.
00:04:57.439 --> 00:04:58.480
Aaron Powell I've seen those demos.
00:04:58.560 --> 00:04:59.279
They look like magic.
00:04:59.360 --> 00:05:00.319
Aaron Powell They do.
00:05:00.639 --> 00:05:07.040
But when we look at the reality of mission-critical systems, that illusion shatters, doesn't it?
00:05:07.120 --> 00:05:09.120
Aaron Powell It shatters on contact with reality.
00:05:09.360 --> 00:05:12.560
I mean, let's look back over the last 40 years of enterprise IT.
00:05:12.639 --> 00:05:15.439
Aaron Powell The universal constraint was always execution.
00:05:15.600 --> 00:05:16.079
Trevor Burrus Right.
00:05:16.319 --> 00:05:20.800
If you needed a new billing system or a healthcare triage app, what did you have to do?
00:05:20.879 --> 00:05:29.199
Aaron Ross Powell You had to secure a massive budget, hire an army of engineers, sit them in a room for two years, and just pay them to manually type out syntax.
00:05:29.279 --> 00:05:30.560
Aaron Powell Line by single line.
00:05:30.959 --> 00:05:32.160
Execution was expensive.
00:05:32.240 --> 00:05:32.879
It was slow.
00:05:32.959 --> 00:05:36.399
It was the primary reason software projects failed or went over budget.
00:05:36.480 --> 00:05:39.040
Aaron Powell Or was just canceled before they ever even saw production.
00:05:39.360 --> 00:05:39.759
Exactly.
00:05:40.000 --> 00:05:45.519
But generative AI has essentially collapsed the cost and time of raw execution to zero.
00:05:45.920 --> 00:05:47.199
It's basically free now.
00:05:47.600 --> 00:05:53.120
Right now, as we speak, any competitor of yours can generate a million lines of code in an afternoon.
00:05:53.519 --> 00:05:56.959
The sheer volume of output is no longer a competitive advantage.
00:05:57.199 --> 00:05:58.480
It is a baseline commodity.
00:05:58.800 --> 00:06:00.319
Generation is commodity.
00:06:00.560 --> 00:06:02.959
Verification is the scarce resource.
00:06:03.360 --> 00:06:04.240
Precisely.
00:06:04.560 --> 00:06:11.040
If anyone can generate a million lines of code, the winner in the enterprise space is no longer the fastest builder.
00:06:11.600 --> 00:06:17.120
The winner is the enterprise that can be trusted to deliver that code securely, compliantly, and accountably.
00:06:17.279 --> 00:06:23.199
Aaron Powell And the market actually has a new term for what happens when you prioritize generation over verification.
00:06:23.600 --> 00:06:23.920
Yes.
00:06:24.319 --> 00:06:25.600
It's called AI slop.
00:06:26.079 --> 00:06:26.959
AI slop.
00:06:27.279 --> 00:06:31.360
We see this popping up everywhere, but what does it actually look like inside a corporate network?
00:06:31.439 --> 00:06:34.639
Aaron Powell AI Slop is basically high volume, low signal output.
00:06:34.720 --> 00:06:36.240
Aaron Powell Okay, so it looks like real code.
00:06:36.319 --> 00:06:39.120
Aaron Ross Powell Oh, it looks completely plausible at first glance.
00:06:39.199 --> 00:06:42.319
It compiles perfectly, it reads like a professional senior engineer wrote it.
00:06:42.639 --> 00:06:43.360
But beneath the surface.
00:06:43.759 --> 00:06:45.040
Structurally flawed.
00:06:45.199 --> 00:06:47.439
It misses the nuanced business context.
00:06:47.600 --> 00:06:53.680
It introduces subtle race conditions, or it's just entirely valueless to the specific needs of the enterprise.
00:06:54.079 --> 00:06:56.079
So it provides the illusion of rigorous work.
00:06:56.160 --> 00:07:00.079
Aaron Powell And that illusion is the true danger of casual AI adoption.
00:07:00.319 --> 00:07:03.439
Before AI, you at least had to spend time to be wrong.
00:07:03.600 --> 00:07:05.279
Now you can be wrong instantly.
00:07:05.360 --> 00:07:06.639
Aaron Ross Powell That is terrifying.
00:07:06.879 --> 00:07:10.720
You can flood your enterprise code base with architectural debt in a matter of seconds.
00:07:10.800 --> 00:07:11.519
Aaron Ross Powell Instantly.
00:07:11.680 --> 00:07:14.560
Let's ground this in that banking scenario you mentioned at the top of the show.
00:07:14.879 --> 00:07:15.199
Okay, yeah.
00:07:15.600 --> 00:07:20.319
Imagine a major Tier 1 bank decides to accelerate their development.
00:07:20.560 --> 00:07:24.240
They equip their engineering teams with unstructured AI coding assistance.
00:07:24.720 --> 00:07:25.839
The engineers are probably thrilled.
00:07:26.240 --> 00:07:27.040
Oh, they love it.
00:07:27.199 --> 00:07:31.839
They ship a new credit decision engine that was co-written with AI.
00:07:32.000 --> 00:07:34.959
It gets deployed faster than any project in the bank's history.
00:07:35.040 --> 00:07:35.199
Trevor Burrus, Jr.
00:07:35.360 --> 00:07:38.639
The CIO is presenting slides to the board, everyone is celebrating.
00:07:38.720 --> 00:07:38.879
Trevor Burrus, Jr.
00:07:39.040 --> 00:07:39.199
Right.
00:07:39.439 --> 00:07:42.000
Until that regulatory audit hits six months later.
00:07:42.399 --> 00:07:43.120
There's always an audit.
00:07:43.439 --> 00:07:43.680
Always.
00:07:43.920 --> 00:07:53.439
The regulatory body comes in, they look at the decision engine and they ask: who decided on this specific risk rule that is currently denying credit to this demographic?
00:07:53.519 --> 00:07:54.079
Aaron Powell Oof.
00:07:54.240 --> 00:07:55.759
On what date was it decided?
00:07:55.839 --> 00:07:59.360
And what was the rigorous business and mathematical justification for it?
00:07:59.439 --> 00:08:02.879
Aaron Powell And the bank has absolutely no answer.
00:08:03.199 --> 00:08:09.759
The engineers look at the repository and say, well, the AI generated that specific module during a chat session.
00:08:09.839 --> 00:08:13.680
Aaron Ross Powell We reviewed it, it compiled, passed the unit tests, so we pushed it.
00:08:13.759 --> 00:08:14.720
Aaron Powell Exactly.
00:08:14.959 --> 00:08:22.000
And in a regulated industry like finance, healthcare, or aerospace, that answer invites catastrophic fines.
00:08:22.079 --> 00:08:22.240
Trevor Burrus, Jr.
00:08:22.480 --> 00:08:24.480
Or think about a public sector procurement team.
00:08:24.639 --> 00:08:28.959
They're under massive pressure to modernize, you know, aging government infrastructure.
00:08:29.120 --> 00:08:29.279
Trevor Burrus, Jr.
00:08:29.439 --> 00:08:30.000
Huge pressure.
00:08:30.079 --> 00:08:34.240
Aaron Ross Powell So they pay a premium to a vendor who promises agentic delivery.
00:08:34.480 --> 00:08:37.919
They buy the expensive copilot licenses, they watch the flashy demos.
00:08:38.080 --> 00:08:40.720
Aaron Powell Where the AI builds a citizen portal in five minutes.
00:08:41.039 --> 00:08:41.200
Right.
00:08:41.519 --> 00:08:46.080
But when the system goes into production, they discover that nothing was structurally re-engineered.
00:08:46.399 --> 00:08:48.559
The underlying legacy processes are exactly the same.
00:08:48.639 --> 00:08:50.799
Trevor Burrus The data silos are identical.
00:08:50.960 --> 00:08:52.559
The governance is completely missing.
00:08:52.960 --> 00:08:55.840
They've just paid millions of dollars for the appearance of modernization.
00:08:56.240 --> 00:09:03.919
While inheriting a tangled mess of undocumented, AI-generated code that literally no human in the building actually understands.
00:09:04.240 --> 00:09:05.120
It's a disaster.
00:09:05.519 --> 00:09:06.799
I have to stop you there, though.
00:09:06.960 --> 00:09:13.679
Because if I'm a CTO listening right now, I'm probably thinking, look, you two are just advocating for traditionalist red tape.
00:09:14.080 --> 00:09:15.120
The slowdown argument.
00:09:15.360 --> 00:09:15.600
Yeah.
00:09:15.840 --> 00:09:18.559
I am in a dogfight with agile startups.
00:09:18.720 --> 00:09:27.120
If I slow down my teams to build all this heavy governance, the startup down the street using ungoverned AI is gonna beat me to market by six months.
00:09:27.600 --> 00:09:27.759
Right.
00:09:28.000 --> 00:09:29.120
Speed is survival.
00:09:29.519 --> 00:09:30.080
Exactly.
00:09:30.240 --> 00:09:34.159
Aren't you just demanding we go back to the dark ages of waterfall development?
00:09:34.240 --> 00:09:37.679
Aaron Powell And that is the exact fear driving the market right now?
00:09:38.000 --> 00:09:39.759
But it's based on a false premise.
00:09:39.840 --> 00:09:40.559
Aaron Powell How so?
00:09:41.039 --> 00:09:44.480
Because speed without control is just industrializing error.
00:09:44.720 --> 00:09:47.200
The fear of moving too slow is valid, sure.
00:09:47.279 --> 00:09:48.720
Aaron Powell But the threats on the other side.
00:09:49.120 --> 00:09:50.080
Are existential.
00:09:50.159 --> 00:09:56.799
The total loss of control, massive regulatory liability, catastrophic data breaches, absolute vendor lock-in.
00:09:57.200 --> 00:09:59.440
Those will kill a company faster than a slow release cycle.
00:09:59.840 --> 00:10:00.080
Absolutely.
00:10:00.320 --> 00:10:04.720
The reframe here is that governance and trust are not speed bumps slowing you down.
00:10:04.960 --> 00:10:06.879
Trust is now the sellable property.
00:10:07.279 --> 00:10:09.759
Governance is the ultimate commercial differentiator.
00:10:10.080 --> 00:10:10.399
Right.
00:10:10.559 --> 00:10:22.480
If your agile competitor ships a product six months faster, but fails their Dora compliance audit or suffers a data breach because their ungoverned AI hallucinated an insecure API endpoint.
00:10:22.559 --> 00:10:24.720
Aaron Powell Their speed didn't buy the market share.
00:10:25.120 --> 00:10:25.360
No.
00:10:25.600 --> 00:10:27.279
It bought them a corporate crisis.
00:10:27.679 --> 00:10:28.080
Wow.
00:10:28.240 --> 00:10:38.480
So the enterprise that wins over the next decade is the one that can prove to its board, its clients, and regulators that it has absolute mathematical control over its AI delivery pipeline.
00:10:38.799 --> 00:10:39.120
Exactly.
00:10:39.279 --> 00:10:46.480
So if execution is genuinely no longer the expensive bottleneck, if the raw manual act of writing the syntax is practically free.
00:10:46.960 --> 00:10:52.879
Then the fundamental economics of what an enterprise decides to build versus what it decides to buy must drastically change.
00:10:53.120 --> 00:10:56.000
It completely inverts the economic model of enterprise IT.
00:10:56.399 --> 00:10:58.000
Let's break that down because this is huge.
00:10:58.399 --> 00:11:00.000
We have to look at the old logic first.
00:11:00.240 --> 00:11:07.679
For decades, the ironclad rule of thumb was that you only built custom software for your strictly core, unique, differentiating business functions.
00:11:08.080 --> 00:11:10.000
The classic make versus buy decision.
00:11:10.320 --> 00:11:10.559
Right.
00:11:10.720 --> 00:11:16.639
And you almost always chose buy for absolutely everything else HR systems, ERPs, finance procurement.
00:11:17.200 --> 00:11:20.080
You bought commodity off-the-shelf software packages.
00:11:20.480 --> 00:11:20.639
Yes.
00:11:20.879 --> 00:11:22.960
And the reason was purely economic.
00:11:23.120 --> 00:11:26.799
Building custom software from scratch took years and cost millions of dollars.
00:11:27.120 --> 00:11:30.799
You had to maintain an army of developers in perpetuity just to keep it running.
00:11:31.200 --> 00:11:44.559
So you bought massive packages like SAP, Oracle, or Salesforce, and you endured the excruciating organizational pain of adapting your unique business processes to fit their generic data models.
00:11:44.879 --> 00:11:45.600
It was a compromise.
00:11:45.840 --> 00:11:46.080
Yeah.
00:11:46.399 --> 00:11:48.399
Painful, but cheaper than building from scratch.
00:11:48.480 --> 00:11:52.639
Aaron Powell But with a gentic engineering, that economic equation is destroyed.
00:11:52.720 --> 00:11:56.320
Aaron Powell Which introduces a new model we need to talk about service as a software.
00:11:56.639 --> 00:11:56.960
Yes.
00:11:57.440 --> 00:11:58.559
Service as a software.
00:11:58.639 --> 00:11:59.279
Let's compare it.
00:11:59.440 --> 00:12:03.600
In the traditional Sauce model software as a service, you subscribe to a generic product.
00:12:03.679 --> 00:12:08.000
Aaron Powell And you mold your human business to fit the software's rigid constraints.
00:12:08.080 --> 00:12:08.559
Aaron Powell Exactly.
00:12:08.879 --> 00:12:11.360
In service as a software, it is the exact inverse.
00:12:11.600 --> 00:12:15.440
The enterprise defines the bespoke exact service they actually need.
00:12:15.519 --> 00:12:18.480
Aaron Powell And the software is custom built around that specific service.
00:12:18.559 --> 00:12:19.120
Aaron Powell Right.
00:12:19.279 --> 00:12:27.759
And crucially, it's delivered in an OpEx model, an operational expenditure model, with no massive upfront capital build cost.
00:12:27.840 --> 00:12:32.320
Aaron Powell That is a revolutionary shift because the cost of execution has plummeted.
00:12:32.480 --> 00:12:38.720
You can now build a massive custom enterprise system in months at a fraction of the historical cost.
00:12:38.799 --> 00:12:40.639
Aaron Powell Think about a specialty insurer.
00:12:40.799 --> 00:12:47.679
For years, they may have been forcing their highly specialized claims adjudicators to use a clunky off-the-shelf tool.
00:12:48.000 --> 00:12:51.600
A tool that didn't fit their unique data models or their regulatory environment.
00:12:52.000 --> 00:12:52.960
Or their company culture.
00:12:53.120 --> 00:12:58.720
They just lived with the friction because building a custom claims engine was, you know, a$20 million proposition.
00:12:58.799 --> 00:13:06.080
Aaron Powell But now, using an agenic practice, they can commission a bespoke claims workflow that perfectly maps to their actual processes.
00:13:06.399 --> 00:13:10.240
And integrates flawlessly with their legacy mainframes, quickly and economically.
00:13:10.639 --> 00:13:12.240
Or imagine a regional hospital network.
00:13:12.399 --> 00:13:12.480
Yeah.
00:13:12.879 --> 00:13:15.679
Need a highly specific triage and bed management system.
00:13:16.000 --> 00:13:21.360
Something that integrates with their unique local clinical pathways and regional ambulance dispatch.
00:13:21.759 --> 00:13:26.639
Previously, commissioning a custom build for that would have been a five-year impossibility.
00:13:32.879 --> 00:13:37.440
Literally risking lives because the software couldn't accommodate their specific word structures.
00:13:37.840 --> 00:13:46.720
But today, they can commission a highly bespoke agent-built system and have it deployed and fully audited in months.
00:13:47.039 --> 00:13:49.200
So the make zone is massively expanded.
00:13:49.600 --> 00:13:50.080
Massively.
00:13:50.240 --> 00:13:54.720
Strategic differentiation no longer belongs only to your absolute core functions.
00:13:54.879 --> 00:13:56.480
It extends everywhere.
00:13:56.879 --> 00:13:59.679
Your HR system can perfectly match your culture.
00:13:59.840 --> 00:14:03.120
Your procurement system can perfectly match your vendor risk profiles.
00:14:03.440 --> 00:14:05.519
It's like the economics of bespoke tailoring.
00:14:05.840 --> 00:14:06.879
I love this analogy.
00:14:07.039 --> 00:14:15.600
For the first time in history, getting a suit custom tailored to your exact, unique measurements costs the exact same amount as buying an ill-fitting one off the rack.
00:14:15.919 --> 00:14:16.080
Right.
00:14:16.240 --> 00:14:22.399
Why would you ever force your business to wear a generic, uncomfortable software suit when you can have a bespoke one for the same price?
00:14:22.720 --> 00:14:26.159
In the same time frame, without the massive CapEx hit, you wouldn't.
00:14:26.559 --> 00:14:27.759
You absolutely wouldn't.
00:14:27.919 --> 00:14:35.519
And that means the crucial question for CTOs and CIOs is no longer how much faster can we build what we used to build?
00:14:35.919 --> 00:14:37.360
What should they be asking this quarter?
00:14:37.679 --> 00:14:42.720
What should we now consider building that we previously had to rule out because it was too expensive?
00:14:42.799 --> 00:14:43.279
Aaron Powell Okay.
00:14:43.519 --> 00:14:45.360
I understand the economic argument.
00:14:45.679 --> 00:14:53.279
But practically speaking, if I'm an engineering director, my immediate fear is complete structural collapse.
00:14:53.360 --> 00:14:53.919
Aaron Powell Sure.
00:14:54.080 --> 00:14:54.720
Quality control.
00:14:55.039 --> 00:14:55.679
Exactly.
00:14:55.919 --> 00:14:59.440
We've established companies are going to build vastly more custom software.
00:14:59.600 --> 00:15:06.879
Won't churning out all this bespoke software at record speed just result in a mountain of buggy rushed products.
00:15:07.039 --> 00:15:12.639
Aaron Powell The old uh if we generate 10 times more code, aren't we generating 10 times more technical debt arguments?
00:15:12.720 --> 00:15:13.200
Trevor Burrus Right.
00:15:13.279 --> 00:15:15.039
It's the most ingrained belief in all of IT.
00:15:15.200 --> 00:15:15.360
Trevor Burrus, Jr.
00:15:15.440 --> 00:15:18.320
The universal law of project management, the iron triangle.
00:15:18.399 --> 00:15:21.679
Trevor Burrus You can have it faster, better, or cheaper, but never all three.
00:15:21.759 --> 00:15:21.919
Trevor Burrus, Jr.
00:15:22.000 --> 00:15:23.039
Usually the conversation is.
00:15:23.200 --> 00:15:25.679
You can have it faster or you can have it better, but never both.
00:15:25.759 --> 00:15:31.840
Aaron Ross Powell Every unhappy conversation in enterprise software delivery over the last 40 years stems from that exact trade-off.
00:15:32.240 --> 00:15:36.399
But serious agentic engineering claims to dissolve this trade-off completely.
00:15:36.879 --> 00:15:38.639
I need you to explain the mechanics of that.
00:15:39.360 --> 00:15:47.519
How is it actually possible to deliver materially better outcomes, tighter architectural fit, higher maintainability, stronger security?
00:15:47.600 --> 00:15:51.519
Trevor Burrus N D, roughly four times more functionality in less calendar time.
00:15:51.600 --> 00:15:52.240
Aaron Powell Right.
00:15:52.480 --> 00:15:53.440
It sounds like magic.
00:15:53.679 --> 00:15:54.559
It is not magic.
00:15:54.639 --> 00:15:58.399
It is a fundamental structural redistribution of human effort.
00:15:58.799 --> 00:16:00.320
Walk me through that redistribution.
00:16:00.639 --> 00:16:09.120
In the old delivery model, 80% of our project's timeline was consumed by raw coding, debugging syntax errors, fighting with CICD pipelines.
00:16:09.440 --> 00:16:10.159
And the other 20%.
00:16:10.639 --> 00:16:17.120
Only 20% was spent on deep exploration of the business problem, rigorous architectural design, and thorough validation.
00:16:17.519 --> 00:16:22.240
But with a genic engineering, that raw coding phase compresses from months down to seconds.
00:16:22.559 --> 00:16:23.200
Exactly.
00:16:23.440 --> 00:16:28.240
But, and this is the absolute critical part people miss, the total human effort does not disappear.
00:16:28.559 --> 00:16:30.639
It expands massively into the upstream phases.
00:16:30.960 --> 00:16:31.200
Yes.
00:16:31.360 --> 00:16:40.000
The humans spend their time on rigorous architectural design, defining mathematically exact specifications, and deep skeptical validation of the AI's outputs.
00:16:40.399 --> 00:16:46.240
Less AI-assisted coding, more editorial board with a specialist agent at every desk.
00:16:46.559 --> 00:16:52.480
Let's really explore that editorial board analogy because it captures the operational reality perfectly.
00:16:52.799 --> 00:16:55.919
Think about a major Pulitzer-winning newspaper.
00:16:56.240 --> 00:16:56.480
Right.
00:16:56.559 --> 00:16:59.279
An editor-in-chief doesn't write every single story.
00:16:59.600 --> 00:17:02.240
They don't go out to the streets to conduct every interview.
00:17:06.000 --> 00:17:08.319
But they own the coherence of the publication.
00:17:08.480 --> 00:17:11.039
They ensure rigorous factual accuracy.
00:17:11.200 --> 00:17:13.039
They maintain the editorial tone.
00:17:13.519 --> 00:17:18.160
And they hold the ultimate legal and reputational accountability for what goes to print.
00:17:18.240 --> 00:17:22.000
Aaron Powell That is the enterprise developer's job now, scaled across a team of agents.
00:17:22.079 --> 00:17:27.359
Aaron Powell So the agents are the reporters out gathering information, drafting the text, structuring the narratives?
00:17:27.759 --> 00:17:35.920
And the human developer is the editor reviewing the drafts, challenging the underlying assumptions, demanding rewrites, and formally approving the final publication.
00:17:36.079 --> 00:17:43.599
Aaron Powell But what happens when the reporter, the developer agent, brings back a draft that is structurally unsound?
00:17:44.000 --> 00:17:51.200
Say it uh hallucinates an open source library that doesn't exist or introduces a known vulnerability.
00:17:51.279 --> 00:17:51.920
Aaron Powell Exactly.
00:17:52.079 --> 00:17:53.200
What does the human editor do?
00:17:53.279 --> 00:17:55.440
Aaron Powell That's where the rest of the editorial board comes in.
00:17:55.599 --> 00:17:58.160
The human editor doesn't just read the code manually.
00:17:58.480 --> 00:18:00.319
They use the test agent, the fact checker.
00:18:00.720 --> 00:18:00.960
Right.
00:18:01.119 --> 00:18:05.200
They run exhaustive automated test suites against the developer agent's code.
00:18:05.519 --> 00:18:09.680
And if the code fails, the human doesn't manually fix the syntax.
00:18:09.759 --> 00:18:15.920
Aaron Powell No, they kick it back to the developer agent with instructions on why it failed and mandate a rewrite.
00:18:16.160 --> 00:18:19.039
The human is orchestrating a system of validation.
00:18:19.519 --> 00:18:24.160
So to be completely clear, we're not saying that human engineers are working fewer hours.
00:18:24.559 --> 00:18:24.960
Not at all.
00:18:25.119 --> 00:18:30.400
The promise of agentic engineering is not that your engineering team gets to play golf every afternoon.
00:18:30.720 --> 00:18:34.960
The exact same amount of human effort is applied, but it is applied entirely differently.
00:18:35.279 --> 00:18:39.279
The headline is more functionality, better quality, less time.
00:18:39.519 --> 00:18:44.880
It's an elevation of human thought, moving from manual labor to high-level systemic judgment.
00:18:45.279 --> 00:18:50.640
You are paying your highly skilled, incredibly expensive engineers to think, to design, and to govern.
00:18:50.880 --> 00:18:56.000
You're no longer paying them to act as human compilers typing out boilerplate rest APIs.
00:18:56.400 --> 00:19:06.960
But to achieve this reality, you cannot just buy your engineers a conversational AI license like Copilot, drop it in their IDE, and expect this to happen.
00:19:07.279 --> 00:19:07.599
No.
00:19:07.759 --> 00:19:11.759
You need a fundamentally new physical structure for how software is made.
00:19:12.160 --> 00:19:14.640
Let's dive into the physical anatomy of agentic engineering.
00:19:14.960 --> 00:19:18.720
This is where we separate the serious enterprise practices from the vendor marketing height.
00:19:19.039 --> 00:19:24.240
A mature agentic engineering practice is built on a very specific, rigid physical structure.
00:19:24.559 --> 00:19:26.400
We need to examine the three-layer architecture.
00:19:26.799 --> 00:19:27.839
The three-layer architecture.
00:19:27.920 --> 00:19:29.599
It consists of three distinct levels.
00:19:29.759 --> 00:19:31.599
First, the governance layer at the top.
00:19:31.839 --> 00:19:33.440
Second, the AI layer in the middle.
00:19:33.680 --> 00:19:36.480
And third, the deterministic core at the base.
00:19:36.799 --> 00:19:40.640
Walk me through exactly what lives in each of those layers and how they interact.
00:19:40.799 --> 00:19:41.680
Let's start with governance.
00:19:42.000 --> 00:19:45.119
The governance layer is entirely human and entirely authoritative.
00:19:45.440 --> 00:19:47.200
This is where human accountability resides.
00:19:47.519 --> 00:19:47.759
Yes.
00:19:47.920 --> 00:19:55.440
It holds the signed business requirements, the legal and compliance sign-offs, the foundational architectural decisions, and the liability.
00:19:55.599 --> 00:19:58.319
It defines the boundaries of what the system is allowed to do.
00:19:58.720 --> 00:20:02.240
When a regulator asks a question, this layer provides the answer.
00:20:02.319 --> 00:20:05.119
Aaron Powell Okay, so beneath that is the AI layer.
00:20:05.519 --> 00:20:09.519
This is where your probabilistic specialized agents live and work.
00:20:09.920 --> 00:20:12.880
They generate code, draft tests, propose architectures.
00:20:13.279 --> 00:20:14.480
Analyze data, yeah.
00:20:14.720 --> 00:20:18.079
They are incredibly powerful, but they are inherently probabilistic.
00:20:18.160 --> 00:20:23.440
Aaron Powell Meaning they can hallucinate, they can be non-deterministic, and they require strict supervision.
00:20:23.839 --> 00:20:24.319
Exactly.
00:20:24.480 --> 00:20:27.680
And beneath that AI layer is the deterministic core.
00:20:27.759 --> 00:20:27.920
Trevor Burrus, Jr.
00:20:28.079 --> 00:20:29.119
The foundation of the enterprise.
00:20:29.440 --> 00:20:35.279
This is where the money flows, where patient lives are tracked, where regulators audit, and where legacy systems operate.
00:20:35.599 --> 00:20:41.440
And crucially, inside the deterministic core, AI is strictly physically forbidden.
00:20:41.519 --> 00:20:42.720
Aaron Powell Completely forbidden.
00:20:42.960 --> 00:20:47.519
The core consists entirely of fully deterministic, traceable, and explainable code.
00:20:47.599 --> 00:20:50.319
Aaron Powell So the AI layer can write code for the core.
00:20:50.640 --> 00:20:51.039
Yes.
00:20:51.279 --> 00:20:58.400
But once that code is approved and deployed into the core, it executes deterministically without any active AI decision making.
00:20:58.480 --> 00:21:01.279
Aaron Powell The Autonomous Train Network analogy is brilliant for this.
00:21:01.680 --> 00:21:03.039
Oh, I love that analogy.
00:21:03.440 --> 00:21:07.599
Think of the three-layer architecture like a modern, fully automated railway.
00:21:07.920 --> 00:21:14.880
The deterministic tracks, the physical steel rails, and the routing switches represent your deterministic core.
00:21:15.200 --> 00:21:18.240
They dictate exactly where the train can and cannot go.
00:21:18.559 --> 00:21:25.440
The massive AI engine provides the incredible momentum and speed, but it is physically constrained by the tracks.
00:21:25.680 --> 00:21:30.319
And the governance layer is the human conductor sitting in the control room holding the liability.
00:21:30.720 --> 00:21:34.559
You contain the probabilistic elements and you own the accountable elements.
00:21:34.960 --> 00:21:35.839
That is exactly it.
00:21:36.000 --> 00:21:42.640
The AI is the engine, but it cannot jump the deterministic tracks and it cannot override the human conductor's emergency brake.
00:21:42.960 --> 00:21:48.079
But for the AI engine to actually do the work, it needs to interact with the real world of enterprise IT.
00:21:48.400 --> 00:21:50.720
It can't just exist in a vacuum chatting with itself.
00:21:51.119 --> 00:21:51.279
Right.
00:21:51.440 --> 00:21:55.599
I want to explain the mechanism behind how these agents actually execute tasks.
00:21:55.759 --> 00:21:57.759
How does an agent log into Jira?
00:21:58.160 --> 00:21:59.759
How does it read a consonance page?
00:22:00.240 --> 00:22:03.680
This brings us to MCP or the Model Context Protocol.
00:22:04.079 --> 00:22:06.319
How does MCP actually work under the hood?
00:22:06.799 --> 00:22:08.880
This is the critical technical bridge.
00:22:09.119 --> 00:22:13.039
Historically, an LLM has a knowledge cutoff date and exists in isolation.
00:22:13.440 --> 00:22:15.680
It doesn't know what your team talked about on Slack yesterday.
00:22:16.000 --> 00:22:19.839
And it doesn't have access to your private on-premise Git repositories.
00:22:20.160 --> 00:22:30.559
So MCP, the Model Context Protocol, is an open standard that allows your AI agents to securely connect to and work inside the actual tools your teams already use every single day.
00:22:30.880 --> 00:22:31.519
Exactly.
00:22:31.759 --> 00:22:35.359
Without MCP, an AI agent is locked in a chat window.
00:22:35.759 --> 00:22:42.880
If you ask it to help with a Jira ticket, it can only spit out text describing what a Jira ticket should look like and the human has to copy-paste it.
00:22:43.200 --> 00:22:46.559
But with MCP, the agent is securely authenticated into your environment.
00:22:46.960 --> 00:22:51.759
It can make an API call to Confluence, pull down the JSON payload of your product requirements document.
00:22:52.079 --> 00:22:56.160
Parse that text into its context window, use that context to formulate a plan.
00:22:56.559 --> 00:23:05.680
And then use another MCP tool connection to physically open a Jira Epic, break it down into user stories, write the code, and push a commit directly to Bitbucket.
00:23:06.079 --> 00:23:11.680
It treats the agent just like a human colleague with a corporate login, subject to the same role-based access controls.
00:23:12.079 --> 00:23:15.200
This is a massive warning flag for enterprise buyers.
00:23:15.599 --> 00:23:16.640
Huge flag.
00:23:16.880 --> 00:23:20.720
If you ask a vendor, where is the AI in your delivery model?
00:23:20.880 --> 00:23:27.200
And their answer is simply, oh, our developers use Chat GPT, or we bought co-pilot licenses.
00:23:27.599 --> 00:23:29.279
They are not doing agentic engineering.
00:23:29.759 --> 00:23:34.000
No, they are using AI as a localized, isolated typing assistant.
00:23:34.240 --> 00:23:38.480
They lack the three-layer architecture, and they are not deeply integrated via MCP.
00:23:38.559 --> 00:23:38.720
Trevor Burrus, Jr.
00:23:38.799 --> 00:23:42.720
Tool-level adoption yields localized, tool-level results.
00:23:43.039 --> 00:23:47.119
If you want enterprise level transformation, you need architectural level integration.
00:23:47.200 --> 00:23:47.680
Aaron Powell Exactly.
00:23:48.079 --> 00:23:49.839
So we have this beautiful three-layer architecture.
00:23:49.920 --> 00:23:54.319
We have our MCP connectors securely authenticating our agents into Jira and GitHub.
00:23:54.720 --> 00:23:56.000
Now we need to populate the system.
00:23:56.480 --> 00:23:56.640
Right.
00:23:56.799 --> 00:24:01.519
Who is actually doing the work and how do we prevent them from generating that dangerous AI slop?
00:24:01.839 --> 00:24:06.400
We need to talk about the absolute discipline of spec-driven teams versus vibe coding.
00:24:06.720 --> 00:24:09.359
I see teams falling into the vibe coding trap constantly.
00:24:09.680 --> 00:24:10.799
It is rampant right now.
00:24:10.960 --> 00:24:13.200
Vibecoding is the absolute amateur approach.
00:24:13.279 --> 00:24:17.599
Aaron Powell They just open a prompt, chat with the LLM until the code sort of looks right, compile it and pray.
00:24:17.680 --> 00:24:18.160
Aaron Powell Yeah.
00:24:18.240 --> 00:24:31.119
It is a human developer sitting in front of a chat window, typing a vague, unstructured prompt like build me a login screen, and just blindly accepting whatever block of code the LLM spits out.
00:24:31.599 --> 00:24:34.319
It relies entirely on the unstructured vibe of the conversation.
00:24:34.400 --> 00:24:41.920
Aaron Powell There is no rigorous documentation, no architectural planning, no security review before the code is generated.
00:24:42.160 --> 00:24:44.559
It is coding by trial and error at light speed.
00:24:44.720 --> 00:24:49.359
Aaron Powell And what happens when a team vibe codes an entire enterprise application?
00:24:49.440 --> 00:24:51.119
Aaron Powell Complete architectural collapse.
00:24:51.359 --> 00:24:53.839
Because the LLM is just guessing at the context.
00:24:54.000 --> 00:24:57.839
It hallucinates dependencies, it ignores enterprise naming conventions.
00:24:57.920 --> 00:25:02.160
Aaron Powell And it creates a tangled mess of spaghetti code that is impossible to maintain.
00:25:02.240 --> 00:25:06.640
Aaron Powell When a bug appears, no one knows how to fix it because no human actually designed the system.
00:25:06.720 --> 00:25:09.200
Aaron Powell Though spec-driven development is the exact opposite of that.
00:25:09.279 --> 00:25:09.599
Aaron Powell Yes.
00:25:09.839 --> 00:25:20.160
In a spec-driven agential environment, the AI agents must produce and the human experts must formally approve foundational, highly detailed artifacts before a single line of executable code is ever written.
00:25:20.240 --> 00:25:26.000
Aaron Powell Let's break down this specialized agent roster because it's not just one giant AI brain doing everything.
00:25:26.079 --> 00:25:30.240
Aaron Powell No, you mentioned earlier that there's one agent per SDLC role.
00:25:30.319 --> 00:25:33.680
Aaron Powell Walk me through the actual workflow of these specialized agents.
00:25:34.000 --> 00:25:36.319
It is a highly specialized sequential team.
00:25:36.480 --> 00:25:37.759
You start with an analyst agent.
00:25:37.839 --> 00:25:39.680
Aaron Powell And its job is not to write code.
00:25:39.759 --> 00:25:40.400
Aaron Powell Exactly.
00:25:40.559 --> 00:25:49.279
Its job is to ingest the raw business request, conduct requirements analysis, surface logical ambiguities in the prompt, and draft the initial specifications.
00:25:53.440 --> 00:25:57.759
Aaron Powell Which takes those specs and breaks them down into scope, timelines, and dependencies.
00:25:58.160 --> 00:25:59.359
Then you have an architecture agent.
00:25:59.759 --> 00:26:05.119
That reviews the enterprise environment, proposes the technology stack, and drafts the structural design.
00:26:05.519 --> 00:26:06.240
It keeps going.
00:26:06.319 --> 00:26:09.839
You have a UX agent drafting user journeys and component structures.
00:26:10.160 --> 00:26:19.920
You have a Scrum Master agent that uses MCP to log into Jira and shard the architectural plan into highly detailed, actionable user stories.
00:26:20.319 --> 00:26:25.279
Then and only then do you activate a developer agent to actually write the syntax.
00:26:25.599 --> 00:26:30.880
Simultaneously, a test agent is generating automated test scripts based on the original requirements.
00:26:31.200 --> 00:26:38.160
And an auditor agent is constantly updating the documentation in the background to ensure it perfectly matches the actual committed code base.
00:26:38.480 --> 00:26:38.880
Exactly.
00:26:39.039 --> 00:26:43.279
Each of these agents is working in parallel, focusing strictly on their specific domain.
00:26:43.680 --> 00:26:48.400
And the reason you must shard the work across specialized agents is deeply technical, right?
00:26:48.720 --> 00:26:55.759
It has to do with how large language models process information, specifically context windows and attention mechanisms.
00:26:56.160 --> 00:27:03.039
If you try to use a single prompt approach, asking one AI model to act as the analyst, the architect, and the coder all at once.
00:27:03.440 --> 00:27:06.880
The model will suffer from what we call lost in the middle degradation.
00:27:07.200 --> 00:27:09.200
It literally forgets the initial instructions.
00:27:09.599 --> 00:27:09.920
Yes.
00:27:10.319 --> 00:27:14.640
The attention mechanism of the LLM degrades as the context window fills up.
00:27:14.799 --> 00:27:20.720
It will start writing code and completely forget the crucial security constraint you mentioned in paragraph two of the requirements.
00:27:21.200 --> 00:27:28.640
So by sharding the work into specialized agents, you keep the context window for each specific task tight, focused, and highly accurate.
00:27:29.039 --> 00:27:31.279
The analyst agent only thinks about analysis.
00:27:31.519 --> 00:27:35.359
The developer agent only thinks about the specific JIRA ticket it was assigned.
00:27:35.759 --> 00:27:39.920
And these agents communicate with each other through those foundational artifacts.
00:27:40.000 --> 00:27:40.240
Trevor Burrus, Jr.
00:27:40.319 --> 00:27:42.640
The solution requirements document, the SRD.
00:27:43.039 --> 00:27:47.200
The architecture decision record, the ADR, and the test concept.
00:27:47.279 --> 00:27:50.240
Aaron Powell These artifacts form the working memory of the project.
00:27:50.400 --> 00:27:51.680
They are the connective tissue.
00:27:51.759 --> 00:28:04.480
Aaron Powell The analyst agent creates the SRD, the human rigidly reviews and proves it, and then that approved SRD becomes the absolute ground truth prompt for the architecture agent to create the ADR and so on.
00:28:04.559 --> 00:28:05.200
Aaron Powell Exactly.
00:28:05.519 --> 00:28:08.079
But let me ask the ultimate accountability question here.
00:28:08.400 --> 00:28:08.559
Okay.
00:28:08.880 --> 00:28:13.279
We have all these sophisticated agents communicating through artifacts and writing incredible code.
00:28:13.519 --> 00:28:13.680
Right.
00:28:13.920 --> 00:28:18.319
The system goes into production, it handles millions of transactions, and then it breaks.
00:28:18.400 --> 00:28:20.319
Aaron Powell A catastrophic failure in production.
00:28:20.400 --> 00:28:20.559
Trevor Burrus, Jr.
00:28:20.720 --> 00:28:21.839
Revenue is hemorrhaging.
00:28:21.920 --> 00:28:30.079
And the vendor points at the screen and says, Well, the developer agent hallucinated that specific module, and the test agent failed to catch it, so it's the AI's fault.
00:28:30.319 --> 00:28:31.279
Who takes the blame?
00:28:31.599 --> 00:28:36.559
Under no circumstances is blame the AI an acceptable answer in an enterprise context.
00:28:36.720 --> 00:28:37.039
Never.
00:28:37.119 --> 00:28:40.880
Aaron Powell Every single agent in that roster has a named human in the loop.
00:28:41.200 --> 00:28:43.839
The human analyst formally signs off on the SRD.
00:28:44.079 --> 00:28:46.559
The human architect formally approves the ADR.
00:28:46.960 --> 00:28:52.240
The human developer performs a rigorous review of the pull request from the developer agent before merging it.
00:28:52.720 --> 00:28:54.319
The human owns the outcome.
00:28:54.559 --> 00:28:55.279
Always.
00:28:55.599 --> 00:28:59.039
The agent is merely a tool of probabilistic generation.
00:28:59.359 --> 00:29:02.240
The human is the locus of deterministic accountability.
00:29:02.400 --> 00:29:02.559
Trevor Burrus, Jr.
00:29:02.720 --> 00:29:04.000
The human owns the outcome.
00:29:04.160 --> 00:29:04.559
Always.
00:29:04.799 --> 00:29:05.599
I love that phrasing.
00:29:05.680 --> 00:29:05.839
Trevor Burrus, Jr.
00:29:05.920 --> 00:29:06.799
It's non-negotiable.
00:29:06.880 --> 00:29:14.880
Trevor Burrus And if human accountability is non-negotiable, enterprises need a four-proof mathematical way to prove that accountability to regulators and auditors.
00:29:14.960 --> 00:29:16.400
Aaron Powell Which brings us to governance.
00:29:16.799 --> 00:29:24.880
Let's talk about the EU AI Act and the concept of the audit trail by construction, because the regulatory landscape is coming for ungoverned AI and it is coming fast.
00:29:25.279 --> 00:29:28.960
The regulatory landscape is fundamentally changing how we must build software.
00:29:29.200 --> 00:29:32.880
We have to look at the upcoming August 2026 EU AI Act.
00:29:33.279 --> 00:29:36.880
This is not some distant theoretical framework being debated by academics.
00:29:37.279 --> 00:29:41.920
No, this is an incoming reality with massive existential financial implications.
00:29:42.319 --> 00:29:52.400
The European Commission's own impact assessment projects a 17% compliance overhead on AI spending just to meet the documentation and risk management requirements of the Act.
00:29:52.480 --> 00:29:52.640
Trevor Burrus, Jr.
00:29:52.799 --> 00:29:54.319
17% compliance overhead.
00:29:54.480 --> 00:29:54.559
Yeah.
00:29:54.880 --> 00:29:55.839
And for smaller providers.
00:29:55.920 --> 00:29:56.079
Trevor Burrus, Jr.
00:29:56.240 --> 00:30:00.960
Industry estimates suggest the effective burden could eat up 30 to 40 percent of their operating profits.
00:30:01.039 --> 00:30:01.200
Trevor Burrus, Jr.
00:30:01.359 --> 00:30:05.119
The 17% compliance tax just for playing in the AI space.
00:30:05.359 --> 00:30:08.079
And what does an audit under this act actually look like?
00:30:08.400 --> 00:30:10.799
They aren't just going to ask if you used AI.
00:30:10.960 --> 00:30:17.279
They are going to demand to see your Article 10 data governance logs and your Article 17 quality management system.
00:30:17.519 --> 00:30:23.680
Trevor Burrus They will demand to see the exact decision-making process that led to a specific piece of AI-generated logic.
00:30:23.759 --> 00:30:23.920
Trevor Burrus, Jr.
00:30:24.079 --> 00:30:24.319
Exactly.
00:30:24.480 --> 00:30:27.039
And if you are vibe coding, you fail that audit instantly.
00:30:27.119 --> 00:30:29.920
Aaron Powell Because you have no systematic logging of the decision-making process.
00:30:30.000 --> 00:30:30.720
Aaron Powell None at all.
00:30:30.960 --> 00:30:32.960
But it's not just the EU AI Act.
00:30:33.039 --> 00:30:36.559
Look at the 2024 Gerara report, the Digital Operational Resilience Act.
00:30:36.640 --> 00:30:36.799
Trevor Burrus, Jr.
00:30:36.880 --> 00:30:42.960
Their accelerated State of DevOps report showed a shocking 7.2% reduction in software delivery stability globally.
00:30:43.039 --> 00:30:47.759
Aaron Powell Why are systems getting less stable when we have the most advanced coding tools in human history?
00:30:47.839 --> 00:30:50.559
Aaron Powell Because of rapid, ungoverned AI adoption.
00:30:50.960 --> 00:30:52.799
Companies are rushing to use AI.
00:30:52.960 --> 00:30:58.880
Developers are flooding continuous integration pipelines with unreviewed AI-generated pull requests.
00:30:59.200 --> 00:31:05.440
They are bypassing architectural governance, and their production systems are becoming measurably more fragile as a result.
00:31:05.839 --> 00:31:10.319
They are generating technical debt faster than their human QA teams can possibly catch it.
00:31:19.519 --> 00:31:23.599
Like this function calculates the sum above a sum function, it's pointless.
00:31:24.000 --> 00:31:24.160
Right.
00:31:24.319 --> 00:31:29.920
So how does agentic engineering solve for this massive compliance burden and this drop in stability?
00:31:30.079 --> 00:31:32.960
What is an audit trail by construction, technically speaking?
00:31:33.279 --> 00:31:38.319
In traditional software development, creating an audit trail is a miserable year-end cleanup exercise.
00:31:38.640 --> 00:31:48.480
A regulator asks for documentation, and a panic team of engineers spends three weeks retroactively trying to write architecture documents for code they deployed six months ago.
00:31:48.799 --> 00:31:54.400
And barely remember, it is expensive, it is highly inaccurate, and it is universally hated.
00:31:54.799 --> 00:32:04.559
But in a mature agencies, audit trail by construction means that rigorous documentation becomes a natural, automatic, unavoidable byproduct of the workflow itself.
00:32:04.880 --> 00:32:07.680
So it's emitted by the process, not assembles after the fact.
00:32:08.000 --> 00:32:08.480
Exactly.
00:32:08.720 --> 00:32:10.799
Explain the mechanism of how that is emitted.
00:32:11.119 --> 00:32:20.319
Because you are using a spec-driven process with specialized agents connected via MCP, every single step leaves an immutable digital fingerprint.
00:32:20.640 --> 00:32:26.079
Every business requirement in the SRD is mathematically tied to an automated test case generated by the test agent.
00:32:26.480 --> 00:32:31.839
That test case is linked via a git hash to a specific code commit generated by the developer agent.
00:32:32.160 --> 00:32:35.359
And that commit is tied directly to a production deployment log.
00:32:35.599 --> 00:32:40.640
The audit trail is built automatically, layer by layer, in real time as the software is constructed.
00:32:41.039 --> 00:32:43.519
Ask for an audit on software delivered last quarter.
00:32:43.680 --> 00:32:46.000
Who decided what, when, why?
00:32:46.240 --> 00:32:48.000
In a traditional setup, people panic.
00:32:48.319 --> 00:32:50.799
In a serious agentic practice, the answer is already written.
00:32:51.200 --> 00:32:57.440
It is sitting there in the repository, flawlessly linking the business intent to the final line of code.
00:32:57.759 --> 00:33:03.839
And this rigorous level of governance leads to a critical commercial concept we call the reversibility guarantee.
00:33:04.240 --> 00:33:06.400
This is the ultimate test of vendor trust, isn't it?
00:33:06.720 --> 00:33:07.440
It really is.
00:33:07.680 --> 00:33:19.920
AI-generated code must be so clean, so structurally sound, adhering strictly to enterprise design patterns, and so deeply documented that if the AI models and your vendor disappeared tomorrow.
00:33:20.400 --> 00:33:25.440
Your internal human team could take over that exact code base without having to rewrite it from scratch.
00:33:25.759 --> 00:33:28.559
It cannot be an impenetrable black box of AI slot.
00:33:28.720 --> 00:33:28.880
Trevor Burrus, Jr.
00:33:28.960 --> 00:33:30.559
If it's not reversible, it's a trap.
00:33:30.640 --> 00:33:33.599
It's ultimate vendor lock-in disguised as innovation.
00:33:34.079 --> 00:33:34.480
Exactly.
00:33:34.640 --> 00:33:37.279
Think about the failure modes if you lack this governance.
00:33:37.440 --> 00:33:37.920
What happens?
00:33:38.319 --> 00:33:39.440
You get vendor sprawl.
00:33:39.599 --> 00:33:42.480
Every team buys a different AI tool with different standards.
00:33:42.559 --> 00:33:51.039
Aaron Powell, you get silent quality loss where the code compiles but introduces subtle architectural vulnerabilities that won't manifest until the system is under heavy load.
00:33:51.440 --> 00:33:56.640
You get context loss between chat sessions, where the AI forgets why a foundational decision was made.
00:33:56.960 --> 00:34:03.039
And ultimately, you get massive legal exposure when the system fails, and you can't prove who authored the failing logic.
00:34:03.519 --> 00:34:14.320
So bringing this back to the commercial implications for the CTOs and CIOs listening, how does this structural shift change the way we actually negotiate and pay for software development services?
00:34:14.639 --> 00:34:18.719
It forces a total shift away from the traditional time and materials model.
00:34:19.039 --> 00:34:20.559
For decades you bought hours.
00:34:20.639 --> 00:34:24.239
You paid a developer an hourly rate to sit at a desk and type.
00:34:24.559 --> 00:34:30.719
But if the AI is doing the typing in three seconds, why on earth are you paying a massive blended hourly rate?
00:34:31.039 --> 00:34:31.440
Right.
00:34:31.599 --> 00:34:35.440
The industry is shifting to a model we call time, tokens, and materials.
00:34:35.679 --> 00:34:39.679
Governance is no longer an overhead cost you try to minimize during procurement.
00:34:39.840 --> 00:34:42.159
It is the core commercial value you are purchasing.
00:34:42.480 --> 00:34:48.639
You are buying verified outcomes, guaranteed accountability, and regulatory compliance, not raw manual effort.
00:34:48.960 --> 00:34:49.440
Precisely.
00:34:49.599 --> 00:34:59.199
In a time, tokens, and materials model, the cost of the AI agents, the actual compute tokens used to generate the code, becomes visible, trackable, and transparent.
00:34:59.599 --> 00:35:06.159
You pay for the human architectural judgment and the raw computational effort tied to a guaranteed auditable outcome.
00:35:06.480 --> 00:35:09.119
It completely changes the procurement conversation.
00:35:09.440 --> 00:35:15.599
We have covered the architecture, we have covered the agents, the spec-driven discipline, and the automated audit trails.
00:35:15.840 --> 00:35:18.800
But what happens to the actual human beings in this equation?
00:35:19.119 --> 00:35:19.440
Right.
00:35:19.599 --> 00:35:27.599
What happens to the senior engineer who has spent the last 20 years of their life proudly identifying as a person who writes beautiful code?
00:35:27.920 --> 00:35:30.079
This is a massive psychological shift.
00:35:30.559 --> 00:35:36.320
Let's move to our final major topic, the human element, and the shifts from coder to AI supervisor.
00:35:36.639 --> 00:35:44.239
This is often the most difficult, painful part of the transition for an enterprise because it is a profound psychological identity shift.
00:35:44.559 --> 00:35:47.280
The daily job of a developer is fundamentally changing.
00:35:47.519 --> 00:35:51.599
They are no longer typing syntax into an IDE for eight hours a day.
00:35:51.840 --> 00:36:03.920
Their new operational reality involves reading complex AI output, steering the behavior of the agents, refining the system configurations, and exercising extreme, highly calibrated architectural judgment.
00:36:04.239 --> 00:36:07.039
It requires a completely different cognitive muscle.
00:36:07.280 --> 00:36:13.280
It's the difference between writing a book from scratch and editing a dense technical manuscript.
00:36:13.599 --> 00:36:22.880
Editing requires you to hold the entire architecture of the book in your head and constantly look for inconsistencies, tone issues, logical fallacies, and structural flaws.
00:36:23.199 --> 00:36:24.880
It is arguably much harder work.
00:36:25.119 --> 00:36:27.679
It is incredibly taxing cognitive work.
00:36:27.840 --> 00:36:32.960
And to manage this transition, organizations have to radically rethink their team structures.
00:36:33.280 --> 00:36:40.960
We are seeing the rise of two specific organizational patterns: the evangelist model and the junior-senior pair.
00:36:41.199 --> 00:36:48.719
The evangelist is an experienced, highly respected engineering leader within the organization who has deeply internalized the agentic workflow.
00:36:49.039 --> 00:36:51.119
They don't just mandate changes from the top down.
00:36:51.519 --> 00:36:52.480
They sit with the team.
00:36:52.559 --> 00:36:55.679
They demonstrate practically how to interrogate the AI's output.
00:36:56.000 --> 00:37:00.880
And they set the cultural tone that this is an elevation of the engineer's role, not a threat to their job security.
00:37:01.199 --> 00:37:07.119
I want to explore the junior-senior dynamic because this fundamentally breaks how we have trained software engineers for decades.
00:37:07.440 --> 00:37:19.280
In a traditional setup, a senior developer mentors a junior developer by giving them small, simple coding tasks, fix this bug, write the simple CRU'd endpoint, and reviewing their work.
00:37:19.679 --> 00:37:24.960
But in an agentic model, the AI agent is functionally acting as the senior coder, right?
00:37:25.599 --> 00:37:31.440
The agent has encyclopedic knowledge of every design pattern, every language syntax, every library.
00:37:31.519 --> 00:37:31.679
Trevor Burrus, Jr.
00:37:31.840 --> 00:37:34.719
That is the paradox at the heart of agentic engineering.
00:37:34.880 --> 00:37:41.360
The agent possesses the raw capability and encyclopedic syntactical knowledge of a highly experienced senior engineer.
00:37:41.599 --> 00:37:44.960
Therefore, it must be supervised by a human senior engineer.
00:37:45.199 --> 00:37:57.440
Only a seasoned human expert has the scars, the architectural wisdom, and the deep business context to know if the agent's highly sophisticated, perfectly compiling output is actually correct for the specific constraints of the enterprise.
00:37:57.840 --> 00:38:05.519
A junior developer simply doesn't have the experience to know if the AI is hallucinating a plausible sounding but fundamentally flawed microservices architecture.
00:38:05.920 --> 00:38:07.599
So what happens to the human juniors?
00:38:07.679 --> 00:38:12.320
How do they learn their craft if they aren't writing the boilerplate code and fixing the typos anymore?
00:38:12.480 --> 00:38:14.000
Do we just stop hiring juniors?
00:38:14.400 --> 00:38:17.760
No, you absolutely must still hire them, but you train them differently.
00:38:17.840 --> 00:38:24.320
Aaron Powell They learn by pairing with the agent on tightly bounded tasks under the strict constant review of the human seniors.
00:38:24.480 --> 00:38:27.679
Aaron Powell The Junior acts as the orchestrator or the interface.
00:38:27.840 --> 00:38:36.800
They work with the agent to generate components, and then they sit side by side with the senior human to review exactly why certain AI decisions were accepted or rejected.
00:38:37.199 --> 00:38:49.119
The junior learns architectural judgment and system design much faster this way because they are exposed to complex, finished code structures immediately rather than spending their first two years just fighting syntax errors.
00:38:59.360 --> 00:39:03.760
You know, the sloppy, tired code a developer writes right before the weekend when they just want to go home.
00:39:04.000 --> 00:39:06.639
Doesn't AI write Friday afternoon code all the time?
00:39:06.719 --> 00:39:10.320
Aaron Powell Well, the fundamental advantage of the AI is that it never gets tired.
00:39:10.480 --> 00:39:12.159
It doesn't know what a Friday afternoon is.
00:39:12.480 --> 00:39:17.840
It applies the exact same level of mathematical rigor to the millionth line of code as it did to the first.
00:39:18.239 --> 00:39:21.280
But there is a very real challenge here regarding human trust.
00:39:21.519 --> 00:39:24.800
AI-generated code often looks unfamiliar to a human engineer.
00:39:25.119 --> 00:39:29.199
It might use syntax, libraries, or patterns that the human isn't used to seeing.
00:39:29.360 --> 00:39:32.079
This unfamiliarity instantly breeds suspicion.
00:39:32.400 --> 00:39:34.480
So how do you overcome that suspicion?
00:39:34.639 --> 00:39:42.079
Because if the senior human just rejects everything the AI does because it doesn't look like their personal style, the whole system grinds to a halt.
00:39:42.480 --> 00:39:47.280
You overcome it by measuring output objectively rather than by stylistic habits.
00:39:47.599 --> 00:39:52.559
You cannot evaluate AI code based on whether it looks exactly like the code you would have written.
00:39:52.960 --> 00:40:03.440
You have to measure it against objective automated metrics, maintainability scores, security vulnerability scans, cyclomatic complexity metrics, and test coverage percentages.
00:40:03.840 --> 00:40:16.000
If the AI code is structurally sound, highly secure, perfectly functional, and passes every objective metric, the human supervisor must learn to accept it, even if it lacks their personal stylistic flair.
00:40:16.400 --> 00:40:20.000
Which means training your humans is a massive undertaking.
00:40:20.079 --> 00:40:23.119
It's not just about showing them how to use a new software tool.
00:40:23.280 --> 00:40:31.360
I want to push back on the idea that you can just send your engineering team to a two-hour vendor webinar, buy them lunch, and call them AI supervisors.
00:40:31.679 --> 00:40:32.800
You absolutely cannot.
00:40:33.039 --> 00:40:35.599
Organizations that try that fail immediately.
00:40:35.760 --> 00:40:38.079
We look at training in two distinct tiers.
00:40:38.400 --> 00:40:40.320
Tier one is the basic onboarding.
00:40:40.400 --> 00:40:41.360
It gets you in the door.
00:40:41.599 --> 00:40:47.760
It teaches you the interfaces, how to rate the prompts, how the MCP connections work, the basic workflow of the agents.
00:40:48.000 --> 00:40:50.559
But Tier One does not change ingrained human behavior.
00:40:50.800 --> 00:40:54.719
Tier two is where the real transformation happens, and it is difficult.
00:40:54.960 --> 00:41:02.320
Tier two is hands-on delivery of real production grade code alongside the agent on a real high-stakes project.
00:41:02.559 --> 00:41:12.559
It is the grueling, frustrating work of unlearning the 20-year habit of jumping straight to typing and building the new habit of rigorous specification and deeply skeptical review.
00:41:12.880 --> 00:41:16.639
Tier two is the required price of entry to truly change the human mindset.
00:41:16.960 --> 00:41:19.840
Well, we have covered a massive amount of ground today.
00:41:20.000 --> 00:41:36.719
We've talked about the shift from execution to trust, the expansion of custom software through services of software, the breaking of the speed versus quality trade-off, the absolute necessity of a rigid three-layer architecture, and the deep discipline of spec-driven, governed delivery.
00:41:36.960 --> 00:41:41.760
Let's bring this entire deep analysis down to immediate actionable steps for the listener.
00:41:42.159 --> 00:41:52.320
If you are an enterprise leader charting your course over the next 12 to 18 months, what are the specific hard questions you must take back to your teams and your vendors tomorrow morning?
00:41:52.480 --> 00:41:55.360
We have prepared six CISP prompts for you.
00:41:55.679 --> 00:41:58.320
Prompt number one is about trust and verification.
00:41:58.480 --> 00:41:59.840
You must ask your partner.
00:42:00.239 --> 00:42:11.280
Can they explain exactly how their agents work mechanically, what specific decisions the human supervisors are required to make, and most importantly, who holds the ultimate legal responsibility when a system fails?
00:42:11.599 --> 00:42:17.039
And if they say the AI is responsible, or if they give a vague answer about shared liability, walk away immediately.
00:42:17.199 --> 00:42:19.840
Prompt number two is about make versus buy.
00:42:20.079 --> 00:42:32.239
You need to gather your executive team, look at your IT roadmap, and ask, what critical capabilities do we previously rule out, or outsource to generic sauce packages simply because they were too expensive and slow to build?
00:42:32.559 --> 00:42:37.920
Which of those should we now build bespoke using service as a software and agentic engineering?
00:42:38.239 --> 00:42:41.519
Propt number three is about EU AI Act readiness.
00:42:41.760 --> 00:42:42.639
This is critical.
00:42:43.039 --> 00:42:44.880
Ask your engineering leads today.
00:42:45.119 --> 00:42:54.079
Can we produce a complete unbroken mathematical trace from a business requirement all the way down to a deployed line of code on demand for every system we ship?
00:42:54.400 --> 00:42:59.199
If you cannot produce that audit trail today, you are fundamentally not ready for August 2026.
00:42:59.519 --> 00:43:01.840
Prop number four is architecture clarity.
00:43:02.000 --> 00:43:04.719
Ask your internal teams and your external vendors.
00:43:04.880 --> 00:43:07.599
Where exactly is the AI operating in our delivery model?
00:43:07.840 --> 00:43:13.199
And more importantly, what systems constitute our deterministic core where AI is strictly physically kept out?
00:43:13.440 --> 00:43:15.440
Prop number five is pricing.
00:43:15.679 --> 00:43:21.920
Look at your current vendor contracts and ask, are we still buying arbitrary hours or are we buying guaranteed outcomes?
00:43:22.320 --> 00:43:26.159
Does the contract make the cost of the agent compute tokens visible?
00:43:26.400 --> 00:43:29.760
Or is it hidden inside a bloated blended day rate?
00:43:30.079 --> 00:43:32.719
And finally, prompt number six is about your people.
00:43:32.880 --> 00:43:35.280
Ask your HR and engineering directors.
00:43:35.519 --> 00:43:37.360
Do we have the evangelists in place?
00:43:37.519 --> 00:43:38.559
Have we identified them?
00:43:38.880 --> 00:43:47.599
Do we have the structured junior-senior pairings required to turn agentic engineering from a cool marketing slogan into a lived daily operational practice?
00:43:47.920 --> 00:43:59.920
Those six questions will rapidly expose the gap between organizations that are just playing with AI as a new toy and organizations that are rigorously using AI as an industrial grade operating model.
00:44:00.320 --> 00:44:02.960
So, what does this all mean for the future of your enterprise?
00:44:03.199 --> 00:44:07.679
We've spent this time breaking down the mechanics, the architecture, and the heavy governance required.
00:44:07.840 --> 00:44:10.320
I want to leave you with one final provocative thought.
00:44:10.639 --> 00:44:18.960
When the cost and the complexity of software creation fundamentally collapse to near zero, the biggest risk to your enterprise isn't that you won't be able to build something.
00:44:19.360 --> 00:44:28.639
The biggest risk is waking up and realizing that you've been limiting your company's entire strategic vision based on technological constraints that no longer exist.
00:44:29.039 --> 00:44:34.880
If the friction of software development drops to near zero tomorrow, what is the very first thing you're gonna build?
00:44:35.039 --> 00:44:36.320
Now handing it back to Rafa.
00:44:36.719 --> 00:44:39.679
And that's it for this episode of Digital Pulse.
00:44:39.840 --> 00:44:44.159
Now bringing it back to a human voice to say, thank you for listening.
00:44:44.400 --> 00:44:47.599
Remember to subscribe and follow Digital Pulse.
00:44:47.760 --> 00:44:49.920
Real insight, zero noise.