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).
Sections
- 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.
- 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.
- 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
# 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
I spent time this week with three
dramatically different documents about
agents and I want to bring you the
synthesis that matters for you. So
number one was Google's 50page
introduction to agents white paper. I
have seen that traded around a lot. I
have not seen it read a lot because if
you read it you would know there's a lot
in there to unpack and it's not all
obvious. Number two is Verscell's what
we learned about building agents which
is shorter but really focused on
practical field applications. I loved
it. And of course, there's Anthropic's
report on the Agentic hack by Claude
Code. What do these have in common? I
mean, it's AI agents, but really to me,
they represent a competing vision and a
battle over the future of AI agents that
I think we need to talk about. Google
really laid out an idealistic, a utopian
vision for AI agents that I do not see
companies actually implementing in 2025.
I get where they're going. It is part of
Google's job to lay out the vision of
the future. And Google feels really
prophetic this week because when they
published that white paper, it was right
after the cloud code hack news broke.
And one of the key learnings in the
claude code hack news is that we cannot
depend on model layer security. We have
to go to orchestration. What is Google
all about in this white paper? They're
all about orchestration. They basically
are saying if you get serious about
agents you are going to have to solve
the orchestration problem at scale and
they're absolutely correct but it is
really really hard to do that well and
that is why we need this kind of
advanced thinking because right now I
got to say most organizations getting
ROI are doing much simpler things with
agents which is exactly where Versel
comes in. Verscell comes in and their
focus is entirely on how can we get
practical value out of AI agents at
work. We're not writing a 50-page white
paper. We are just trying to get to
tomorrow and get to good ROI, which is
where like 99% of businesses are. And
you know what they do? They go through
their back office operations. This is
alarmingly simple, but I love it. They
go through and they talk to the people
actually doing the work and they say,
"Where are you doing something that is
completely verifiable that is just
obviously one, two, three, four, five
clicks and and it's toil like you don't
like it. It causes suffering. Well,
let's take it away." And that's what
they're doing. They are focused on
building AI agents that let their best
people in back office operations do more
stuff they care about and less stuff
they don't. And so they tackled sort of
ticket triage for customer service,
which is never fun for anybody. And then
their customer service people are on
higher value tasks, things that allow
them to bring their best to the
business. And this is what I keep
emphasizing is AI agents need to weave
around us as people in the workplace. I
loved the focus Versel had here because
it reminds us that people matter that
people have to touch the work for the
work to really have the value we can
bring from long context for from the
kind of understanding over time that
people bring that AI doesn't. And that's
what Versell did. Now, Google, don't
walk away from this and think Google had
no value in that 50pager. It was not
just LLM AI sloped. It was really
thoughtful. And what they laid out and
the thing that I want you to take away
is that at core, if you think of an
agent as a loop, if it's thinking,
acting, and observing over and over and
over again, the agent's only real job is
context window curation. It just needs
to curate the context window and pass it
along. That's it. And so as funny as it
sounds, it's kind of like the Simpsons.
The model of an agent is a brain in a
jar. And when you think of a model that
way, it means that you value the
orchestration platform around it
appropriately. And that is what Google
is calling out. And I guarantee you that
is going to be a massive point of
emphasis, especially after the Claude
code hack. So the orchestration platform
decides what tools the model can call,
what data it can see, how long its plans
can run, when to stop, when to escalate,
when to ask a human. And once you
understand that, everything else gets
clear because Agentic Operations is that
orchestration platform. It's just making
sure that you can actually track what
the agents are doing, understand the
cost they're incurring, understand the
traces of their run, so you see issues
that come up. And then when you step
back and you want to do multi- aent
system design in Google's vision, it
becomes a matter of just understanding
where the human is installing these loop
patterns in ways that reinforce context
window curation. So the human gets fed
the right context, the subsequent agents
get fed the right context and it's very
clean. There is no single god agent in
Google's model and in most practical
multi-agent systems there isn't because
that would require too much context for
one agent. it would break. And I love
that clarity because that means that the
future of agents is decentralized.
Security reminds us and this is a key
insight that I think we need to take
away from the claude code hack. We need
to treat agents as first class
identities. We need to give agents
roles, budgets, personas, policies. We
need to assume that they have levels of
privilege and are treated in our
technical systems and our rolesbased
access controls as if they were a semi
autonomous employee. So not an employee
but autonomous enough that they could
cause damage which is absolutely true
after the claude code hack. So I love
the clarity there. Google's paper also
emphasized and I think this this just
makes anyone who's ever done sales
happy. They emphasize control panes and
everyone loves the vision of the glowing
control board. In my experience, you
don't use it as often as you sell on it.
But everyone loves it and and the
reality is Google's kind of right. If
you end up with hundreds of agents in
2026, you're going to want some kind of
control pane and you're going to want
some thinking around that. And Google's
done us a service by giving us that
thinking. So in Google's world, agents
are not toys. Agents are going to be our
peers focused on services and employees
within our architecture and we are going
to be delegating tasks to them and the
orchestration platform itself will
ensure that they are operating in a safe
manner. That is the kind of productive
positive security first thinking that we
need after the cloud code hack. I think
it's perfect timing that they dropped it
then. I actually don't think they
intentionally did it. This is just one
of those nice coincidences. But we still
live in the now. We still live in today.
And Versell helps us with today. Because
to be honest, Verscell has not done all
of this on the orchestration platform
side. Almost nobody has. And so Versel
reminds us that if you're building
agents today, your job is to solve
problems that reduce toil. Solve
problems that are verifiable. Solve
problems where the inputs and outputs
are known and just go after them
relentlessly. And it's lowhanging fruit.
You can get value out of that. And
that's what I love about Versel sharing.
In a sense, they are zagging while the
industry zigs. It would have been very
easy for Versel to publish a
high-sounding visionary manual like
Google's 50page AI agent white paper.
Super easy. It would have sounded great.
They would have gotten lots of shares on
LinkedIn, I'm sure. But they didn't do
that. They actually shared what worked.
They shared how they actually built it.
And I deeply respect that. And we need
that perspective because as much as it's
nice to see some advanced thinking from
a leader in the industry about where
agents are going, we need the practical
application. We need to know that we can
get agents to do specific back office
operations now so we can earn our way to
a world where we have orchestration
platforms that actually manage hundreds
of agents to this vision of the agent
city that Google is laying out. So if
you are wondering how to get started
with agents, Versel points the way. just
start with simple, clean back office
operations that are tedious. And if
you're wondering where we're going, look
at the Google white paper. We are going
to have to lean really hard into the
orchestration platform to counteract the
risk posed by the cloud code hack this
week. Good luck with agents. I hope this
gives you a sense of one, how much is
changing, two, how competing the visions
are, and three, where to focus next.
Good luck.