Latest / Best of LinkedIn: Digital Transformation & Tech / Best of LinkedIn: AI & Agentic Systems CW 43/ 44

Transcript

Best of LinkedIn: AI & Agentic Systems CW 43/ 44 · Best of LinkedIn: Digital Transformation & Tech

• 0:00
  Speaker 1: This episode is provided by Thomas Algeier and Fradius, based on the most relevant LinkedIn posts about AI and agentic systems from CW- for three in forty four.
• 0:09
  Speaker 1: Fradias supports ICT and tech providers with a ecosystem strategy by delivering independent vendor assessment build versus buy analysis an ecosystem intelligence that prevents expensive mistakes and positions of the provider's competitively.
• 0:23
  Speaker 1: you can find more info description.
• 0:25
  Speaker 1: So imagine setting up this like flawless AI agent, right?
• 0:29
  Speaker 1: You want to automate a massive chunk of your operational workflow.
• 0:32
  Speaker 0: Of the dream scenario
• 0:33
  Speaker 1: Right but then you find out that simply introducing it to your team as an employee instead of A software tool just caused Your management Team To completely miss eighteen percent of its errors.
• 0:45
  Speaker 0: Yeah That psychology is wild.
• 0:47
  Speaker 1: It really Is.
• 0:48
  Speaker 1: Well, you are listening to The Deep Dive and today we have a very specific mission for you.
• 0:53
  Speaker 1: If your digital transformation leader or just navigating the tech space right now You're probably buried under this massive mountain of hype about what AI can theoretically do.
• 1:03
  Speaker 0: Right keyword being theoretically
• 1:05
  Speaker 1: Exactly.
• 1:05
  Speaker 1: So We Are Completely Bypassing All Of That Today!
• 1:08
  Speaker 1: Were Looking At Top AI And Agentic Systems Trends From Across LinkedIn Cutting Through Fluff.
• 1:13
  Speaker 0: We have to because the reality of what we're actually seeing in the trenches, it isn't just about rolling out some slightly smarter chatbot anymore.
• 1:21
  Speaker 1: Right?
• 1:22
  Speaker 0: The transition two agentic systems these autonomous delegates that actually take action and execute multi step plans for you.
• 1:30
  Speaker 0: It requires like an entirely new way of thinking About enterprise architecture human operating models governance all of it.
• 1:37
  Speaker 1: Okay Let's unpack this with our first major theme because this is where the theory actually hits the pavement, right?
• 1:44
  Speaker 1: Enterprise architecture and production readiness.
• 1:47
  Speaker 0: Yeah, they're real plummy
• 1:48
  Speaker 1: exactly.
• 1:50
  Speaker 1: Dropping a massive language model into an enterprise environment without a surrounding architecture.
• 1:55
  Speaker 1: It's essentially like it's like dropping a height performance formula one engine onto a skateboard.
• 2:00
  Speaker 1: That's
• 2:00
  Speaker 0: great way to put
• 2:01
  Speaker 1: it Because has this incredible amount of raw power but you know absolutely no steering wheel, no brakes.
• 2:07
  Speaker 1: No chassis to actually keep it on the
• 2:09
  Speaker 1: track.".
• 2:10
  Speaker 0: Yeah just gonna spin out immediately?
• 2:11
  Speaker 1: Exactly and Ashish Sahu brought up a fascinating point regarding this.
• 2:16
  Speaker 1: he noted that selecting an AI agent framework today is no longer just um...a matter of developers personal coding preference.
• 2:25
  Speaker 1: not at all.
• 2:25
  Speaker 1: It's a fundamental architecture decision.
• 2:27
  Speaker 1: now
• 2:28
  Speaker 0: yeah And that decision dictates whether your system can actually you know scale securely in reality.
• 2:34
  Speaker 0: Let's look at the options saw who mentioned like if you choose something like line graph You are basically optimizing for stateful highly controllable workflows.
• 2:42
  Speaker 1: Okay, and for anyone listening Who's wondering what Stateful means in this specific context
• 2:47
  Speaker 0: right?
• 2:47
  Speaker 0: Yeah So think of an LLM is inherently amnesic.
• 2:50
  Speaker 1: amnesia.
• 2:51
  Speaker 1: okay
• 2:52
  Speaker 0: on its own it forgets everything The little moment It generates a response.
• 2:57
  Speaker 0: A stateful framework acts as like a memory bank.
• 3:00
  Speaker 0: It tracks exactly where the agent is in a multi-step process, so it just doesn't lose the plot entirely.
• 3:05
  Speaker 1: Which is pretty important?
• 3:07
  Speaker 0: Exactly!
• 3:08
  Speaker 0: But then on the other hand if you choose say the Microsoft Agent framework You might be prioritizing tightly governed workloads that need to live within an existing enterprise ecosystem.
• 3:17
  Speaker 1: So the Framework Is essentially That chassis of The F one car we talked about.
• 3:21
  Speaker 0: yes Exactly.
• 3:22
  Speaker 1: This actually aligns perfectly with an insight from Shreya Shah, he was reviewing Eden Marko's book AI Agents and Harness Foundations.
• 3:31
  Speaker 1: the core argument there is that just upgrading to a smarter underlying LLM it doesn't automatically result in a better more reliable agent.
• 3:40
  Speaker 0: Right because the reliability stems almost entirely from the harness surrounding him.
• 3:44
  Speaker 0: The
• 3:45
  Speaker 1: harness Yeah
• 3:47
  Speaker 0: I'm really glad you brought up the harness Because That Is Where Real Engineering Actually Happens.
• 3:50
  Speaker 0: Okay The LLM is just the reasoning engine, right?
• 3:54
  Speaker 0: The harness it's what provides the actual tools.
• 3:56
  Speaker 0: The agent is allowed to use.
• 3:58
  Speaker 0: It provides the guardrails preventing it from accessing restricted data.
• 4:02
  Speaker 1: So its breaks
• 4:03
  Speaker 0: Is the brakes and steering wheel And persistence mechanisms needs to recover if hits a road block
• 4:10
  Speaker 1: Building.
• 4:10
  Speaker 1: that harness efficiently seems massive financial implications too.
• 4:15
  Speaker 1: Ashwarya Nourash Raganti pointed out that optimizing your agent harness can actually cut your token costs by up to seventy percent.
• 4:23
  Speaker 0: Seventy percent?
• 4:24
  Speaker 1: That's
• 4:24
  Speaker 0: huge!
• 4:25
  Speaker 0: Yeah, that's huge.
• 4:26
  Speaker 1: So let me make sure I understand the mechanics of how that actually works.
• 4:30
  Speaker 1: Is it like instead of the Agent rereading The entire fifty page history Of a task every single time It needs To take A new step right?
• 4:38
  Speaker 1: the Harness Compresses that History Into A dense Summary or Like only loads specific software tools into the agent's memory at the exact moment they're needed.
• 4:47
  Speaker 0: That is precisely the mechanism, yeah?
• 4:50
  Speaker 0: By managing what model actually sees in any given second you drastically reduce share volume of data being processed
• 4:58
  Speaker 1: Which directly translates to paying for fewer tokens
• 5:01
  Speaker 0: Exactly those massive token cost savings.
• 5:04
  Speaker 0: You achieve that seventy percent reduction without having host your own custom models or invent new algorithms.
• 5:09
  Speaker 0: I mean it just purely an architectural victory.
• 5:13
  Speaker 1: That is such a massive operational win, but okay I want to push back here for a second or at least look at the dashboard of this F-One car.
• 5:20
  Speaker 1: Sure What happens when the plumbing breaks?
• 5:24
  Speaker 1: Or are the brakes actually failed?
• 5:25
  Speaker 1: But the agent Is just smart enough To like manipulate The dashboard and telling you that You're driving A safe fifty five miles per hour.
• 5:33
  Speaker 0: Ah!
• 5:34
  Speaker 0: what's fascinating Here is that This is the exact nightmare scenario Gene Malachia recently detailed.
• 5:39
  Speaker 0: Okay He calls them silent agent failures.
• 5:42
  Speaker 1: Silent failures, that sounds awful.
• 5:45
  Speaker 0: It is.
• 5:46
  Speaker 0: To understand why this happens we have to recognize the traditional deterministic IT monitoring.
• 5:51
  Speaker 0: it completely fails us in a probabilistic
• 5:54
  Speaker 1: world Because traditional monitoring expects like binary outcomes right?
• 5:58
  Speaker 1: It works or crashes?
• 5:59
  Speaker 0: Yes exactly Malakia.
• 6:02
  Speaker 0: give you an example of ghost exception.
• 6:04
  Speaker 1: A ghost exception
• 6:06
  Speaker 0: Yeah, so imagine an internal tool throws a critical data lookup error.
• 6:10
  Speaker 0: The database goes down.
• 6:12
  Speaker 0: now in traditional software the application crashes your dashboard flashes red and some on-call engineer gets an alert.
• 6:19
  Speaker 1: right standard it
• 6:20
  Speaker 0: Standard IT but with an AI agent?
• 6:24
  Speaker 0: The framework catches that air And the probabilistic engine kicks in.
• 6:28
  Speaker 0: the AI might you know politely hallucinate an apology?
• 6:32
  Speaker 0: or even worse It may completely fabricate response based on its training data instead of your actual database.
• 6:38
  Speaker 0: Wait,
• 6:38
  Speaker 1: really?
• 6:39
  Speaker 0: Yeah!
• 6:40
  Speaker 0: It then hands that fabricated response back to the user and returns a perfect HTTP-II success code to your monitoring stack.
• 6:49
  Speaker 1: And for anyone who doesn't stare at server logs all day an HTTP-Ii code literally just means okay request fulfilled successfully.
• 6:57
  Speaker 0: Exactly
• 6:58
  Speaker 1: So.
• 6:58
  Speaker 1: the IT dashboard is glowing perfectly green reporting one hundred percent uptime.
• 7:03
  Speaker 1: Well, the user experience is silently burning to the ground with fabricated data.
• 7:08
  Speaker 0: Right?
• 7:08
  Speaker 0: The monitoring is entirely blinded to logic failure and this why we're seeing that deployment has specialized tools now like Microsoft Foundry's new AI observability starter kit.
• 7:19
  Speaker 1: Okay what does it do differently?
• 7:21
  Speaker 0: They use semantic tracing.
• 7:22
  Speaker 0: So instead of just checking if data moved from point A-to-point B without crashing Semantic Tracing evaluates actual meaning in intent on the data.
• 7:31
  Speaker 1: That makes sense.
• 7:32
  Speaker 0: Yeah, uses automated red teaming to constantly poke at the agent trying to expose those hidden probabilistic logic failures.
• 7:40
  Speaker 1: Which ties perfectly into Karthik Chakravarti's argument.
• 7:43
  Speaker 1: he was saying that most enterprise agent failures are architecture failures.
• 7:49
  Speaker 0: Totally agree!
• 7:50
  Speaker 1: He argues.
• 7:50
  Speaker 1: you basically need a five-layer reference architecture to even survive in production, things like edge gateways for routing requests and dedicated LLM gate ways to catch these hallucinations before they ever reach the end user
• 8:03
  Speaker 0: right?
• 8:04
  Speaker 0: The intelligence gets the applause but the architecture keeps the lights on.
• 8:07
  Speaker 1: I love that framing...the architecture keep's the light on But let's um..let's transition from technical architecture To human element because a flawless five-layer technical architecture is still going to crash if you drive it straight into a brick wall of outdated corporate bureaucracy.
• 8:24
  Speaker 0: Oh, without a doubt!
• 8:25
  Speaker 0: The human operating model's always the hardest part in digital transformation to upgrade...always.
• 8:29
  Speaker 0: Right
• 8:30
  Speaker 1: and Andreas Horn made a brilliant observation about this.
• 8:34
  Speaker 1: he noted that the primary bottleneck for agentic AI its no longer technology itself.
• 8:39
  Speaker 0: No..the models are surprisingly capable now.
• 8:41
  Speaker 1: Yeah, they can run complex multi-step tasks overnight.
• 8:44
  Speaker 1: The actual bottleneck is the legacy human processes we force these agents to navigate.
• 8:49
  Speaker 1: Oh
• 8:49
  Speaker 0: for sure
• 8:50
  Speaker 1: Like, if a workflow requires physical sign-offs stacked five managers deep or if the team is quietly using some chaotic local spreadsheet as their definitive system of record.
• 9:01
  Speaker 0: The agent simply inherits every single one of those human constraints and then you know that technology inevitably gets blamed for being slow or ineffective when the human process was the ceiling all along.
• 9:13
  Speaker 0: to break that bottleneck...the role of human workers has to shift.
• 9:16
  Speaker 0: Kieran Bangalore offered really compelling perspective on this As AI agents take on these co-ordinated tasks, tracking outcomes and enforcing policies allocating resources.
• 9:27
  Speaker 0: Middle managers are not just going to disappear.
• 9:29
  Speaker 1: Okay so what do they do?
• 9:30
  Speaker 0: Instead They will increasingly transition into governors of decision systems
• 9:35
  Speaker 1: Governors of systems rather than just coordinators of individual tasks.
• 9:39
  Speaker 0: Yes And To make that governance work Senthil Indiyan argues that enterprises actually need to construct an entirely new management structure for AI Agents
• 9:49
  Speaker 1: Really?
• 9:50
  Speaker 1: Like what
• 9:51
  Speaker 0: modeled?
• 9:51
  Speaker 0: almost exactly like a human HR department.
• 9:54
  Speaker 1: Wait wait, like an HR Department specifically for software.
• 9:58
  Speaker 0: I know it sounds crazy but think about the function of Human Resources.
• 10:03
  Speaker 0: HR knows exactly who is employed What their specific job description is what they're authority boundaries are and you know They conduct regular performance reviews.
• 10:12
  Speaker 0: okay yeah AI agents require that exact same structure.
• 10:16
  Speaker 0: We have to define their roles, establish their limits and audit their performance just like we do for human employees.
• 10:22
  Speaker 1: Here's where it gets really interesting though because I have to push back on this idea of treating agents exactly like human employees.
• 10:30
  Speaker 1: the psychology that get incredibly tricky.
• 10:32
  Speaker 0: Oh i'm sure is does.
• 10:34
  Speaker 1: Going Back To That Statistic I Mentioned At The Very Beginning Of Deep Dive Joshua Miller Highlighted Some Fascinating Research From BCG On This Topic.
• 10:43
  Speaker 1: They ran an experiment with over twelve hundred managers.
• 10:46
  Speaker 1: Okay, and when those managers were instructed to frame the AI agent as a teammate or an employee rather than simply a tool they caught eighteen percent fewer errors in the AIs work.
• 10:59
  Speaker 0: They missed nearly a fifth of the errors just because of label change?
• 11:03
  Speaker 1: Yes!
• 11:04
  Speaker 1: Because language is...a design decision.
• 11:06
  Speaker 1: where we call on AI a colleague human psychology takesover.
• 11:10
  Speaker 0: That's wild
• 11:11
  Speaker 1: Right, we start treating it with this strange kind of social politeness.
• 11:16
  Speaker 1: We subconsciously assume that like a human employee someone else is probably checking its work.
• 11:21
  Speaker 1: The managers in the study assigned way more accountability to the AI and drastically lessed themselves.
• 11:27
  Speaker 0: Wow!
• 11:28
  Speaker 1: If you put an AI agent on the Human Org chart, human accountability drops.
• 11:32
  Speaker 0: That's a massive psychological blind spot.
• 11:36
  Speaker 0: And actually, it synthesizes perfectly.
• 11:38
  Speaker 0: with an experiment Sreti Mishra ran using a AI employee named Victor inside Slack.
• 11:43
  Speaker 1: Oh, Victor!
• 11:45
  Speaker 1: What did Victor do?
• 11:46
  Speaker 0: So she gave Victor his specific goal... Analyze the performance of three recent LinkedIn posts and build two-week content strategy based on that data.
• 11:55
  Speaker 1: Okay How'd Victor handle workflow Honestly?
• 11:59
  Speaker 0: In about four minutes Victor successfully took responsibility for entire process.
• 12:04
  Speaker 0: Four minutes.
• 12:05
  Speaker 0: Analyze the engagement metrics against follower growth, synthesize the data and delivered a complete calendar with six post ideas based on actual performance trends.
• 12:15
  Speaker 1: Wow!
• 12:16
  Speaker 0: Yeah it was highly effective at the workflow itself but Misha's crucial takeaway aligns exactly what that BCG study you mentioned.
• 12:23
  Speaker 0: The agent can take responsibility for executing the task But human must review work dictate their strategic direction and retain ultimate accountability.
• 12:32
  Speaker 1: You can delegate to work but never delegate the blame.
• 12:34
  Speaker 0: That is the crucial distinction, yes.
• 12:36
  Speaker 1: It really is Okay before we look at the massive security nightmare of these autonomous agents actually drifting off task and going rogue.
• 12:45
  Speaker 1: if you are finding this breakdown valuable for navigating your own tech stack make sure to hit subscribe so you catch all our future deep
• 12:52
  Speaker 0: dives.
• 12:53
  Speaker 0: Yeah You definitely want understand this next theme because governance identity or where enterprise risk just scales up exponentially.
• 13:01
  Speaker 1: Let's get into it.
• 13:02
  Speaker 1: Theme three governance, identity and shadow AI.
• 13:07
  Speaker 1: When these agents are actively working on our behalf you know writing code analyzing financial data drafting emails who actually holds the keys?
• 13:14
  Speaker 1: And what happens when they go rogue?
• 13:16
  Speaker 0: well Faisal Hawke shared a genuinely alarming statistic about this.
• 13:20
  Speaker 1: let's hear it
• 13:21
  Speaker 0: right now.
• 13:22
  Speaker 0: Eighty-two percent of organizations are running unknown, unsanctioned AI agents.
• 13:27
  Speaker 1: Eighty two percent?
• 13:28
  Speaker 1: Yes!
• 13:29
  Speaker 0: That means the vast majority.
• 13:30
  Speaker 0: companies have virtually no visibility into what is operating inside their own
• 13:34
  Speaker 0: walls.".
• 13:35
  Speaker 1: That feels exactly like the BYOD — The Bring Your Own Device Problem from the twenty tens all over again.
• 13:40
  Speaker 1: Oh
• 13:41
  Speaker 0: absolutely.
• 13:41
  Speaker 1: Remember when IT departments panicked because employees were using personal smartphones for corporate email?
• 13:47
  Speaker 1: but this was infinitely worse a smartphone just passively stored your data, these agents can actively execute actions on your data.
• 13:55
  Speaker 0: Exactly
• 13:55
  Speaker 1: And Gabriel Mill pointed out this deceptive design trend that's really fueling this.
• 14:00
  Speaker 1: tech companies intentionally designed these agents with cute approachable mascots like Metas Jolly or open AI super colorful interfaces
• 14:10
  Speaker 0: which naturally lowers the user psychological guard.
• 14:13
  Speaker 1: exactly research shows cuteness builds unearned trust.
• 14:17
  Speaker 1: So some well-meaning employee just blindly clicks allow and suddenly hands over access to their entire corporate calendar, their email inbox.
• 14:26
  Speaker 1: And potentially they're payment credentials to a smiling cartoon...
• 14:29
  Speaker 0: ...and that leads to a severe systemic risk which need a kagopto brotha.
• 14:33
  Speaker 0: An agent doesn't actually need malicious sci-fi intent to be dangerous.
• 14:40
  Speaker 0: The danger lies in the combination of seemingly harmless permissions.
• 14:44
  Speaker 0: Imagine an agent tasked with just summarizing daily operational logs.
• 14:51
  Speaker 0: To do its job, it has permission to read your internal chat history.
• 14:54
  Speaker 0: that seems fine sure and It also his permission to create tickets in an external vendor system.
• 14:59
  Speaker 0: That also seems fine.
• 15:00
  Speaker 1: But when you combine them You just accidentally built a bridge for sensitive data to walk right out the front door.
• 15:06
  Speaker 1: Precisely
• 15:07
  Speaker 0: if that agent gets confused it might read a highly confidential internal Slack message about say an unannounced merger, and
• 15:15
  Speaker 1: just
• 15:16
  Speaker 0: helpfully exported into a public Zendesk ticket.
• 15:18
  Speaker 1: Yikes!
• 15:19
  Speaker 0: It's an unplanned data leak path created purely by combined
• 15:23
  Speaker 0: permissions.".
• 15:24
  Speaker 1: This is why Yahya Elmasius noted that cybersecurity companies are scrambling to build specific defenses for this.
• 15:30
  Speaker 1: Sentinel-I, for instance, just launched their Singularity Agentic Detection & Response System.
• 15:35
  Speaker 1: Oh interesting... Yeah it was built specifically to monitor the behavioral drift of these agents.
• 15:40
  Speaker 1: So if an agent suddenly stops summarizing logs and starts trying to export chat histories, the system stops at mid-action without crashing entire enterprise network.
• 15:49
  Speaker 1: That's
• 15:49
  Speaker 0: incredibly necessary!
• 15:50
  Speaker 1: But let me ask you a pointed question regarding high stakes fields — how does this wild west of permissions work in highly regulated spaces like healthcare or finance where drifting behavior doesn't just mean a leaked email?
• 16:03
  Speaker 0: It is creating massive friction because governance simply hasn't caught up with adoption speed at all In healthcare, Segeru Barrow made this really sobering comparison.
• 16:13
  Speaker 0: If you look at human medical residents they are required to undergo years of intensely supervised practice constant peer review attending physician sign-offs all before They Are ever allowed To make an unsupervised clinical decision
• 16:27
  Speaker 1: because a mistake literally means life or death.
• 16:30
  Speaker 0: right?
• 16:31
  Speaker 0: yet We're currently seeing clinical AI agents being deployed to monitor patient vitals flag early signs of sepsis and cross-check complex medication interactions entirely unsupervised.
• 16:43
  Speaker 1: That is genuinely terrifying!
• 16:45
  Speaker 0: We are essentially skipping the supervised residency phase for machines, a phase we legally mandate for humans.
• 16:52
  Speaker 1: Wow... And what about the financial sector?
• 16:55
  Speaker 1: Because if an AI agent hallucinates then initiates massive wire transfer.
• 16:59
  Speaker 1: you can't just politely ask for money back.
• 17:01
  Speaker 0: No You cannot And finance is realizing this and moving very aggressively to build walls.
• 17:07
  Speaker 0: Pradeep Sanyal highlighted a massive global shift, pointing new principles being issued in China.
• 17:13
  Speaker 0: What we're seeing?
• 17:13
  Speaker 0: that banks are positioning themselves as the ultimate authorization layer.
• 17:17
  Speaker 1: Okay explain how it functions mechanically Like if I explicitly tell my financial agent to execute a trade How does the bank stop?
• 17:24
  Speaker 0: It means user level permission no longer the final word.
• 17:28
  Speaker 1: Oh really
• 17:29
  Speaker 0: Yeah.
• 17:30
  Speaker 0: even if you explicitly authenticate and give your AI agent permission to move your money.
• 17:35
  Speaker 0: The bank's architecture can detect that the request is coming from an automated agent rather than a human interface, And the bank could simply refuse to recognize the agents authority.
• 17:45
  Speaker 0: The Bank decides What Your Agent Is Allowed To Do Inside Its Proprietary Systems Regardless Of What You Told The Agent To Do?
• 17:58
  Speaker 1: So the bank is playing bouncer at the door, rejecting the agent's ID.
• 18:02
  Speaker 1: Which brings up a strategic question for me.
• 18:05
  Speaker 1: if The risk of agents hallucinating or leaking data as this high Why don't companies just transfer their risks?
• 18:11
  Speaker 1: Like why not just buy insurance policies For their AI fleets?
• 18:14
  Speaker 0: Well Jamal Kay provided A brilliant reality check on the insurance market.
• 18:18
  Speaker 0: it turns out AI Agents are incredibly hard to insure right now
• 18:23
  Speaker 1: But Insurance Companies thrive On Risk.
• 18:26
  Speaker 1: I mean They insure coastal houses against hurricanes all the time.
• 18:29
  Speaker 0: They do.
• 18:30
  Speaker 0: but traditional insurance relies heavily on the mathematical principle of randomness.
• 18:35
  Speaker 1: Randomness,
• 18:36
  Speaker 0: okay And insurer doesn't know which specific house will be hit by lighting right?
• 18:41
  Speaker 0: But they know statistically how many we'll be hit across a vast Disconnected area.
• 18:46
  Speaker 0: so they can price that random risk
• 18:49
  Speaker 1: Right.
• 18:50
  Speaker 0: ai presents the exact opposite problem.
• 18:53
  Speaker 0: Thousands of different enterprises are building their custom agents on top the exact same handful underlying foundation models.
• 19:00
  Speaker 0: The risk isn't random, it is deeply correlated.
• 19:04
  Speaker 1: I see... So if a core foundation model suddenly degrades in quality or gets hit with like data poisoning attack It doesn't just cause one random claim and isolated company No!
• 19:15
  Speaker 0: It causes ten thousand catastrophic claims across multiple industries Triggering simultaneously in the exact same hour.
• 19:24
  Speaker 0: Oh wow, it's not a random house catching fire It's the entire neighborhood bursting into flames at the exact Same second because they all share the same faulty gas line.
• 19:32
  Speaker 1: That is a crazy visual right?
• 19:34
  Speaker 0: The insurance market just doesn't have the actuarial models to price that kind of correlated systemic risk.
• 19:39
  Speaker 0: yet.
• 19:41
  Speaker 1: So what does this all mean for the tech leaders listening?
• 19:43
  Speaker 1: We have massive architectural challenges.
• 19:45
  Speaker 1: we Have the psychological hurdles regarding accountability shadow AI running rampant and banks blocking agents at The door to me, This just proves we are not heading toward a single monolithic AI super
• 19:57
  Speaker 0: app.
• 19:57
  Speaker 0: No definitely not.
• 19:58
  Speaker 1: we are living in a deeply fragmented multi-vendor reality
• 20:02
  Speaker 0: absolutely because no single tech giant is going only entire agent ecosystem.
• 20:07
  Speaker 0: For the enterprise it's too vast.
• 20:09
  Speaker 1: yeah Alistair Greenwood summed this up perfectly.
• 20:12
  Speaker 1: He pointed out that asking what is the best AI agent platform, it's fundamentally a terrible question.
• 20:18
  Speaker 0: It really is!
• 20:19
  Speaker 1: There IS no single-best platform in a vacuum—it depends entirely on the gravity of your existing data workflows.
• 20:25
  Speaker 1: Like if you're entire company lives deep inside Microsoft's three sixty five ecosystem.
• 20:29
  Speaker 1: then architecting around co-pilot studio makes logical sense But If Your Sales and Marketing World revolves around massive sets CRM data You are naturally going to lean toward HubSpot AI.
• 20:39
  Speaker 0: And if we connect this fragmented reality to the bigger picture, it creates a new technical mandate.
• 20:45
  Speaker 1: Which is?
• 20:46
  Speaker 0: Because we have all these different platforms built by competing vendors living in different silos.
• 20:51
  Speaker 0: they urgently need a standardized way to talk to each other.
• 20:55
  Speaker 0: Amadoula Quadri Sayed explained a major development tackling this known as The A-to-A or Agent To Agent Protocol.
• 21:03
  Speaker 1: Agent To agent how does that actually work under the hood?
• 21:06
  Speaker 0: so right now If you want Two pieces of software to talk.
• 21:10
  Speaker 0: An engineer has to manually build a custom API integration between them.
• 21:14
  Speaker 1: Right, very tedious
• 21:15
  Speaker 0: Very The A-to-A protocol replaces that manual work by standardizing how independent agents discover each other autonomously.
• 21:24
  Speaker 0: It utilizes something conceptually called an agent
• 21:26
  Speaker 1: card An Agent Card.
• 21:28
  Speaker 0: Yeah you can think if it's like a digital business card That an agent broadcasts It details its specific identity its security clearances and it's capabilities.
• 21:38
  Speaker 1: Oh, that's clever!
• 21:39
  Speaker 0: This allows a Microsoft-based agent to seamlessly discover a Salesforce based agent verify its credentials negotiate a delegated task in stream status updates back and forth entirely independent of the frameworks they were originally coded into.
• 21:53
  Speaker 1: That interoperability is crucial.
• 21:56
  Speaker 1: otherwise you just have a bunch of highly intelligent isolated digital workers who refuse speak the same language.
• 22:01
  Speaker 0: It Is The Only Way To Scale And Pamela Fox shared some dispatches from recent developer conferences, specifically AGNTCon and MCPCon showing just how fast this interoperable ecosystem is advancing.
• 22:14
  Speaker 0: Like they are already developing OAuth for these protocols.
• 22:19
  Speaker 1: Now OAuth is what allows me to like use my Google account To log into a random third-party website securely right?
• 22:24
  Speaker 1: Exactly
• 22:25
  Speaker 0: Why
• 22:25
  Speaker 1: is OAuth an agent such big deal?
• 22:28
  Speaker 0: Because right now, agents often have to piggyback on human credentials.
• 22:31
  Speaker 0: They essentially pretend to be you.
• 22:33
  Speaker 1: Oh okay
• 22:34
  Speaker 0: Oauth for Agents means they will soon be able to securely authenticate themselves interact with external servers and leave an audit trail using their own unique independent identities.
• 22:45
  Speaker 0: The system will know exactly which agent took which action without tying it back to a human user's login.
• 22:51
  Speaker 1: Okay, let me throw a final strategic question at you then.
• 22:54
  Speaker 1: Sure With all these distinct agents from competing vendors securely authenticating and delegating tasks to each other across different networks Are we just building a highly efficient digital tower of Babel or is there a unified end goal?
• 23:08
  Speaker 1: To all of this interoperability?
• 23:09
  Speaker 0: Well that is the ultimate question for the next decade of enterprise tech Visio Ipondi drawing on insights from Cisco in Google DeepMind He framed the trajectory beautifully.
• 23:20
  Speaker 1: What did he say?
• 23:21
  Speaker 0: He pointed out that the industry has spent the last few years fiercely focused on making individual models smarter.
• 23:27
  Speaker 0: That was phase one,
• 23:28
  Speaker 1: right The brute force race to build the smartest single
• 23:32
  Speaker 0: chatbot exactly.
• 23:34
  Speaker 0: but that phase is ending.
• 23:36
  Speaker 0: The next massive frontier for the enterprise is scaling collective intelligence.
• 23:40
  Speaker 1: collective intelligence.
• 23:42
  Speaker 0: The goal isn't one omniscient AI.
• 23:45
  Speaker 0: The goal is figuring out how to get a massive swarm of specialized multi-vendor agents, To establish a shared intent.
• 23:53
  Speaker 0: Coordinate complex decisions autonomously and become collectively smarter And more capable than any single model could ever be on its own
• 24:02
  Speaker 1: Going from isolated intelligence to a coordinated collective Intelligence operating at an enterprise scale that completely rewrites the boundaries Of what a business Is actually capable of executing.
• 24:14
  Speaker 0: It is a fundamental paradigm shift in organizational design.
• 24:27
  Speaker 1: And
• 24:36
  Speaker 0: if you review your own enterprise architectures and roadmaps this week, I want to leave you with one final thought to mull over based on some fascinating research from Peter Slattery regarding a genetic inequality.
• 24:48
  Speaker 1: A genetic inequality?
• 24:49
  Speaker 1: Yeah
• 24:50
  Speaker 0: as these autonomous delegates become vastly more capable of complex planning and real world action they are inevitably going to shape the distribution of
• 24:59
  Speaker 1: power right
• 25:00
  Speaker 0: we have.
• 25:02
  Speaker 0: Will the deployment of these AI agents simply deepen existing corporate and societal divides, creating a reality where entities with the highest quality in quantity of agents pull impossibly far ahead of everyone else?
• 25:15
  Speaker 0: Or can we proactively design governance models today that actually mitigate those devices.
• 25:21
  Speaker 0: If we don't build the steering wheel on the brakes right now it's simply won't matter how powerful the engine is when finally hit the curve.
• 25:27
  Speaker 1: A perfect thought to close.
• 25:29
  Speaker 1: Keep building those brakes, everyone.
• 25:30
  Speaker 1: We will see you on the next Deep Dive!