Learning Library

← Back to Library

AI Agents: Vision Versus Practice

Key Points

  • Google’s 50‑page white paper sketches a utopian, orchestration‑centric vision for AI agents that many companies aren’t yet able to implement, especially after the Claude‑code hack showed model‑level security is insufficient.
  • The Anthropic “Agentic hack” report underscores that reliable AI agents must rely on robust orchestration rather than trusting the model itself for security.
  • Versel’s shorter, field‑focused guide argues that most businesses achieve real ROI by deploying simple, low‑complexity agents that automate repetitive back‑office tasks like ticket triage.
  • Versel’s practical approach emphasizes augmenting human workers—freeing staff from tedious chores so they can focus on higher‑value activities—contrasting with Google’s more speculative roadmap.
  • Together, the three documents illustrate a clash between a lofty, future‑focused vision (Google), a security‑driven reality check (Anthropic), and a near‑term, ROI‑driven implementation strategy (Versel).

Full Transcript

# AI Agents: Vision Versus Practice **Source:** [https://www.youtube.com/watch?v=LNpp73qHbJA](https://www.youtube.com/watch?v=LNpp73qHbJA) **Duration:** 00:08:22 ## Summary - Google’s 50‑page white paper sketches a utopian, orchestration‑centric vision for AI agents that many companies aren’t yet able to implement, especially after the Claude‑code hack showed model‑level security is insufficient. - The Anthropic “Agentic hack” report underscores that reliable AI agents must rely on robust orchestration rather than trusting the model itself for security. - Versel’s shorter, field‑focused guide argues that most businesses achieve real ROI by deploying simple, low‑complexity agents that automate repetitive back‑office tasks like ticket triage. - Versel’s practical approach emphasizes augmenting human workers—freeing staff from tedious chores so they can focus on higher‑value activities—contrasting with Google’s more speculative roadmap. - Together, the three documents illustrate a clash between a lofty, future‑focused vision (Google), a security‑driven reality check (Anthropic), and a near‑term, ROI‑driven implementation strategy (Versel). ## Sections - [00:00:00](https://www.youtube.com/watch?v=LNpp73qHbJA&t=0s) **Competing Visions for AI Agents** - A synthesis of three recent AI‑agent papers shows a clash between Google's utopian orchestration roadmap, Verscell’s pragmatic field guide, and Anthropic’s security‑focused hack analysis. - [00:03:48](https://www.youtube.com/watch?v=LNpp73qHbJA&t=228s) **Decentralized Agent Orchestration Platforms** - The speaker explains that AI agents function like “brains in jars” managed by an orchestration layer that controls tool access, context flow, human interaction, and cost tracking, emphasizing a decentralized, multi‑agent architecture highlighted by recent security concerns such as the Claude code hack. - [00:07:05](https://www.youtube.com/watch?v=LNpp73qHbJA&t=425s) **Versel’s Pragmatic AI Agent Approach** - The speaker lauds Versel for sharing concrete, back‑office AI agent implementations instead of lofty vision papers, positioning them as a practical starter while urging developers to watch Google’s broader orchestration roadmap. ## Full Transcript
0:00I spent time this week with three 0:01dramatically different documents about 0:04agents and I want to bring you the 0:06synthesis that matters for you. So 0:09number one was Google's 50page 0:10introduction to agents white paper. I 0:12have seen that traded around a lot. I 0:14have not seen it read a lot because if 0:15you read it you would know there's a lot 0:17in there to unpack and it's not all 0:18obvious. Number two is Verscell's what 0:21we learned about building agents which 0:23is shorter but really focused on 0:25practical field applications. I loved 0:27it. And of course, there's Anthropic's 0:29report on the Agentic hack by Claude 0:31Code. What do these have in common? I 0:34mean, it's AI agents, but really to me, 0:36they represent a competing vision and a 0:39battle over the future of AI agents that 0:41I think we need to talk about. Google 0:43really laid out an idealistic, a utopian 0:47vision for AI agents that I do not see 0:50companies actually implementing in 2025. 0:53I get where they're going. It is part of 0:55Google's job to lay out the vision of 0:57the future. And Google feels really 1:00prophetic this week because when they 1:02published that white paper, it was right 1:04after the cloud code hack news broke. 1:07And one of the key learnings in the 1:09claude code hack news is that we cannot 1:11depend on model layer security. We have 1:13to go to orchestration. What is Google 1:15all about in this white paper? They're 1:17all about orchestration. They basically 1:20are saying if you get serious about 1:22agents you are going to have to solve 1:24the orchestration problem at scale and 1:26they're absolutely correct but it is 1:29really really hard to do that well and 1:31that is why we need this kind of 1:33advanced thinking because right now I 1:35got to say most organizations getting 1:37ROI are doing much simpler things with 1:39agents which is exactly where Versel 1:41comes in. Verscell comes in and their 1:44focus is entirely on how can we get 1:47practical value out of AI agents at 1:49work. We're not writing a 50-page white 1:50paper. We are just trying to get to 1:53tomorrow and get to good ROI, which is 1:55where like 99% of businesses are. And 1:58you know what they do? They go through 2:00their back office operations. This is 2:03alarmingly simple, but I love it. They 2:05go through and they talk to the people 2:06actually doing the work and they say, 2:08"Where are you doing something that is 2:11completely verifiable that is just 2:13obviously one, two, three, four, five 2:15clicks and and it's toil like you don't 2:18like it. It causes suffering. Well, 2:21let's take it away." And that's what 2:22they're doing. They are focused on 2:25building AI agents that let their best 2:28people in back office operations do more 2:32stuff they care about and less stuff 2:34they don't. And so they tackled sort of 2:36ticket triage for customer service, 2:38which is never fun for anybody. And then 2:40their customer service people are on 2:42higher value tasks, things that allow 2:45them to bring their best to the 2:46business. And this is what I keep 2:47emphasizing is AI agents need to weave 2:50around us as people in the workplace. I 2:53loved the focus Versel had here because 2:55it reminds us that people matter that 2:59people have to touch the work for the 3:01work to really have the value we can 3:04bring from long context for from the 3:06kind of understanding over time that 3:09people bring that AI doesn't. And that's 3:11what Versell did. Now, Google, don't 3:13walk away from this and think Google had 3:15no value in that 50pager. It was not 3:17just LLM AI sloped. It was really 3:19thoughtful. And what they laid out and 3:22the thing that I want you to take away 3:23is that at core, if you think of an 3:26agent as a loop, if it's thinking, 3:28acting, and observing over and over and 3:30over again, the agent's only real job is 3:33context window curation. It just needs 3:37to curate the context window and pass it 3:40along. That's it. And so as funny as it 3:43sounds, it's kind of like the Simpsons. 3:45The model of an agent is a brain in a 3:48jar. And when you think of a model that 3:50way, it means that you value the 3:52orchestration platform around it 3:55appropriately. And that is what Google 3:57is calling out. And I guarantee you that 3:58is going to be a massive point of 4:00emphasis, especially after the Claude 4:02code hack. So the orchestration platform 4:04decides what tools the model can call, 4:07what data it can see, how long its plans 4:09can run, when to stop, when to escalate, 4:11when to ask a human. And once you 4:13understand that, everything else gets 4:15clear because Agentic Operations is that 4:18orchestration platform. It's just making 4:20sure that you can actually track what 4:23the agents are doing, understand the 4:25cost they're incurring, understand the 4:27traces of their run, so you see issues 4:28that come up. And then when you step 4:30back and you want to do multi- aent 4:32system design in Google's vision, it 4:34becomes a matter of just understanding 4:37where the human is installing these loop 4:40patterns in ways that reinforce context 4:42window curation. So the human gets fed 4:44the right context, the subsequent agents 4:46get fed the right context and it's very 4:47clean. There is no single god agent in 4:50Google's model and in most practical 4:52multi-agent systems there isn't because 4:55that would require too much context for 4:57one agent. it would break. And I love 4:59that clarity because that means that the 5:01future of agents is decentralized. 5:04Security reminds us and this is a key 5:07insight that I think we need to take 5:08away from the claude code hack. We need 5:10to treat agents as first class 5:13identities. We need to give agents 5:15roles, budgets, personas, policies. We 5:19need to assume that they have levels of 5:23privilege and are treated in our 5:24technical systems and our rolesbased 5:26access controls as if they were a semi 5:29autonomous employee. So not an employee 5:32but autonomous enough that they could 5:34cause damage which is absolutely true 5:37after the claude code hack. So I love 5:38the clarity there. Google's paper also 5:40emphasized and I think this this just 5:42makes anyone who's ever done sales 5:44happy. They emphasize control panes and 5:47everyone loves the vision of the glowing 5:48control board. In my experience, you 5:50don't use it as often as you sell on it. 5:52But everyone loves it and and the 5:54reality is Google's kind of right. If 5:56you end up with hundreds of agents in 5:572026, you're going to want some kind of 6:00control pane and you're going to want 6:01some thinking around that. And Google's 6:03done us a service by giving us that 6:05thinking. So in Google's world, agents 6:07are not toys. Agents are going to be our 6:09peers focused on services and employees 6:13within our architecture and we are going 6:16to be delegating tasks to them and the 6:19orchestration platform itself will 6:21ensure that they are operating in a safe 6:24manner. That is the kind of productive 6:27positive security first thinking that we 6:29need after the cloud code hack. I think 6:31it's perfect timing that they dropped it 6:33then. I actually don't think they 6:34intentionally did it. This is just one 6:35of those nice coincidences. But we still 6:38live in the now. We still live in today. 6:40And Versell helps us with today. Because 6:43to be honest, Verscell has not done all 6:45of this on the orchestration platform 6:46side. Almost nobody has. And so Versel 6:49reminds us that if you're building 6:51agents today, your job is to solve 6:53problems that reduce toil. Solve 6:55problems that are verifiable. Solve 6:57problems where the inputs and outputs 6:58are known and just go after them 7:00relentlessly. And it's lowhanging fruit. 7:03You can get value out of that. And 7:05that's what I love about Versel sharing. 7:07In a sense, they are zagging while the 7:09industry zigs. It would have been very 7:10easy for Versel to publish a 7:13high-sounding visionary manual like 7:15Google's 50page AI agent white paper. 7:18Super easy. It would have sounded great. 7:20They would have gotten lots of shares on 7:21LinkedIn, I'm sure. But they didn't do 7:23that. They actually shared what worked. 7:25They shared how they actually built it. 7:26And I deeply respect that. And we need 7:28that perspective because as much as it's 7:30nice to see some advanced thinking from 7:33a leader in the industry about where 7:34agents are going, we need the practical 7:36application. We need to know that we can 7:39get agents to do specific back office 7:41operations now so we can earn our way to 7:44a world where we have orchestration 7:46platforms that actually manage hundreds 7:48of agents to this vision of the agent 7:51city that Google is laying out. So if 7:54you are wondering how to get started 7:55with agents, Versel points the way. just 7:58start with simple, clean back office 8:00operations that are tedious. And if 8:02you're wondering where we're going, look 8:03at the Google white paper. We are going 8:05to have to lean really hard into the 8:07orchestration platform to counteract the 8:08risk posed by the cloud code hack this 8:11week. Good luck with agents. I hope this 8:13gives you a sense of one, how much is 8:16changing, two, how competing the visions 8:18are, and three, where to focus next. 8:20Good luck.