Six Principles for Enterprise AI Agents
Key Points
- AI agents are already production‑ready at Fortune 100 firms like Walmart, which has automated 95% of its bug fixes with 200 specialized agents, so waiting years to adopt them is a costly mistake.
- The first principle for successful deployment is “architecture first”: build a model‑agnostic orchestration layer that manages and swaps specialized agents, because the architecture (not the specific model) provides lasting competitive advantage.
- Investing early in AI agents yields “learning compounds,” where early adoption accelerates capability growth and creates a market‑leading feedback loop over 18‑24 months.
- Effective orchestration involves breaking workflows into discrete tasks, assigning appropriate agents, aggregating results, and applying guardrails and governance to ensure reliability and flexibility.
- Major tech players (e.g., Microsoft) are consolidating around unified agent frameworks, indicating that 80% of enterprises are moving toward orchestration layers rather than direct model integration.
Sections
- Act Now on AI Agents - The speaker urges companies to stop postponing AI agents, citing Walmart's success and outlining six production-ready principles, beginning with an architecture‑first, model‑agnostic approach.
- Deploy Memory‑Augmented AI Agents Now - The speaker urges organizations to implement memory‑augmented AI agents within an orchestration layer that learn contextual workflows—citing Walmart bug‑triage and JP Morgan use‑case deployments—to capture strategic value immediately without waiting.
- Vertical Defensibility Through Specialized Workflows - The speaker argues that embedding domain‑specific data, rules, and orchestration around an LLM creates a defensible vertical solution that remains valuable—and is even enhanced—when more powerful generic models appear.
- Strategic Agent Deployment for Velocity - The speaker outlines how introducing AI agents into high‑impact workflows—using a six‑step decision cascade from architecture and memory to automated vertical processes—cuts context‑switching costs and delivers compounding speed‑up dividends across teams.
Full Transcript
# Six Principles for Enterprise AI Agents **Source:** [https://www.youtube.com/watch?v=-oI7mrudRn8](https://www.youtube.com/watch?v=-oI7mrudRn8) **Duration:** 00:17:25 ## Summary - AI agents are already production‑ready at Fortune 100 firms like Walmart, which has automated 95% of its bug fixes with 200 specialized agents, so waiting years to adopt them is a costly mistake. - The first principle for successful deployment is “architecture first”: build a model‑agnostic orchestration layer that manages and swaps specialized agents, because the architecture (not the specific model) provides lasting competitive advantage. - Investing early in AI agents yields “learning compounds,” where early adoption accelerates capability growth and creates a market‑leading feedback loop over 18‑24 months. - Effective orchestration involves breaking workflows into discrete tasks, assigning appropriate agents, aggregating results, and applying guardrails and governance to ensure reliability and flexibility. - Major tech players (e.g., Microsoft) are consolidating around unified agent frameworks, indicating that 80% of enterprises are moving toward orchestration layers rather than direct model integration. ## Sections - [00:00:00](https://www.youtube.com/watch?v=-oI7mrudRn8&t=0s) **Act Now on AI Agents** - The speaker urges companies to stop postponing AI agents, citing Walmart's success and outlining six production-ready principles, beginning with an architecture‑first, model‑agnostic approach. - [00:04:20](https://www.youtube.com/watch?v=-oI7mrudRn8&t=260s) **Deploy Memory‑Augmented AI Agents Now** - The speaker urges organizations to implement memory‑augmented AI agents within an orchestration layer that learn contextual workflows—citing Walmart bug‑triage and JP Morgan use‑case deployments—to capture strategic value immediately without waiting. - [00:07:53](https://www.youtube.com/watch?v=-oI7mrudRn8&t=473s) **Vertical Defensibility Through Specialized Workflows** - The speaker argues that embedding domain‑specific data, rules, and orchestration around an LLM creates a defensible vertical solution that remains valuable—and is even enhanced—when more powerful generic models appear. - [00:13:48](https://www.youtube.com/watch?v=-oI7mrudRn8&t=828s) **Strategic Agent Deployment for Velocity** - The speaker outlines how introducing AI agents into high‑impact workflows—using a six‑step decision cascade from architecture and memory to automated vertical processes—cuts context‑switching costs and delivers compounding speed‑up dividends across teams. ## Full Transcript
leaders. If anyone in your org is
saying, "We're not ready for AI agents,"
they're wrong. If you are asking, "Are
we ready for AI agents? Are AI agents
mature enough for us?" You're wrong.
While you are figuring out how to
perfect your AI readiness deck, Walmart
has automated 95% of their bug fixes
with 200 specialized agents. AI agents
are here in production at Fortune 100
companies right now. Here are the six
principles that Walmart and others have
used to build AI agents into production
workflows. And you can absolutely go
after that this quarter. You do not have
to wait. This is not a 2026
conversation. One of the things that I
worry about is that companies are going
to take AI agents too slowly like
traditional software. Like, okay, maybe
we'll be ready. We'll add it in 2026.
No. If you see an AI agent use case, go
after it and go after it aggressively.
So, what are those six principles? We're
going to dive right in. Principle number
one is architecture first thinking. So,
executives tend to ask me, Nate, which
model should we bet on for AI agents?
Wrong question, guys. The right question
is which architecture delivers the
workflows we're looking for? Model
advantage lasts maybe a quarter at best.
Architectural advantage is something
that persists for years. So the
principle is this. You want to build
model agnostic orchestration that isn't
dependent on particular models to
execute particular workflows. In
particular, you want to make sure that
your agents are selected, evaluated,
swapped, and combined individually. And
you do that based on the architecture
and what it demands, the workflows,
right? This comes back to your
orchestration layer being the
competitive bet, not the models. The
models are replaceable commodities. you
can build a model orchestration layer
which is exactly what Walmart built and
that is going to get you much much
farther. So, you want to be in a place
where you can deploy central
orchestrators that manage specialized
agents like Walmart's YB model is what
they use. WIBY, you can look it up. They
are designed to tackle a range of tough
tasks with appropriate tooling, right?
You can go after that by separating out
your workflows into tasks, delegating
those appropriately, aggregating them
up, having appropriate guardrails and
governance. It's system design. You're
designing a system in which models sit
and do particular tasks with particular
tools. That way you will have
infrastructure that enables you to
switch models regardless of what comes
out tomorrow. Microsoft cares so much
about this they retired their autogen
framework entirely to unify around an
agent framework. 80% of enterprises out
there are starting to get into
orchestration layers roughly speaking.
If you were going directly to models,
you're building AI agents on sand. It's
not going to last. Principle number two,
learning compounds. And with AI agents,
you need to invest early in order to get
ahead of the market on compounded
learning in 18 to 24 months. JP Morgan
has had 18 months of institutional
learning and you cannot buy, steal or or
replicate that with better tech. They
have spent 18 months on AI agents
already in late 2025. Every day you
wait, that gap widens. Right now, you
may not be competing with JP Morgan, so
you don't care. But that's what we're
talking about when we talk about
compounding learning. Memory systems
need to be turned on early to turn
adoption into a permanent advantage for
you. Organizational learning
accumulates. competitors are going to
have to face unclosable gaps because you
are going to have agents trained on more
data than they have always. Always. So
you want to be in a place where you can
deploy early and capture significant
percentage gains in accuracy in
completion etc. You basically preserve
an accuracy and quality gain as an edge
over time because you are always getting
the best models against a growing body
of data that other people don't have.
This gives you asymmetric strategic
value. So you want to implement memory
augmented agents inside an orchestration
layer that learn your organizational
context that track the information that
matters when it matters and how it
relates to decisions. And yes, the
examples I'm giving you, that's exactly
what's happening in task appropriate
ways. In this case, for Walmart, it was
triaging bug tickets. That was really
important. That's what they focused on.
Focus on systems that are able to absorb
and understand meaningful workflows.
Make sure they have clear terminology
they can learn from, like dictionaries.
Make sure they have clear workflow
patterns. to make sure they have really
really crisp precedents to work against
how things have worked in the past, how
you've done b business logic in the
past, how you handle data ambiguities,
and make sure they have exception paths.
200,000 JP Morgan employees have gotten
450 use cases into AI agent land.
Basically, they've they've uploaded
those use cases. They've had AI agents
start to get to get into those use
cases. A system deployed today will
understand your organization as well as
JP Morgan does now for those 450 use
cases in 2027. Right? Don't wait. That's
the whole point of this video. Don't
wait. AI agents are ready now. That is
one of the big things that has shifted
in the past few months is that there is
no reason to wait. It isn't even as
expensive as it used to be. AI agent
SDKs are out there for major model
makers all over the place. They're out
there for cloud providers. You can
assemble AI agent frameworks. Your team
can. Principle number three. Let's talk
about workflows. I've been referencing
it a little bit. Demos are going to get
you applause or get your engineers
applaud, but but the workflows that I'm
talking about are really aimed at real
ROI. And that is the difference that
most executives talk about when they get
AI right. Enterprise AI spending is
scaling by like triple digits a year in
percentage point terms. But it's scaling
because boring automation prints return
on investment. It really like the the
return on investment is skyhigh if you
can pick specific automations for
defined workflows and stick agents
against them. And so you should be
looking not at feature counts, not at
the number of agents you launch, not at
login metrics, not at the number of
tickets even. You should be looking at
the rate of correct completion for
workflows that matter to the business.
Can you get to 90% plus AI automated
completion for workflows that matter to
the business? Recurring automation will
compound in value in a way that flashy
demos will never do. If you can get to,
you know, call it 30% automation in
month one and you have a dedicated team
working to drive that number, the value
just compounds. is that team bites off
more edge cases and is able to knock out
that workflow and they go on to the next
workflow and eventually it gets easier
and easier as you start to stamp these
out. Identify the highfrequency
high-cost workflows first. Of course,
you want to focus on workflows that have
enough defined inputs that you can have
tools called against defined inputs and
the agent can make correct decisions at
a very high rate to fully complete the
workflow. If you can't get that, pick a
different workflow initially because if
you pick a high ambiguity workflow
initially, it's going to be hard to get
success. You'll grow discouraged and you
probably won't go after workflows with
the aggression that you need to.
Principle number four, vertical
defensibility.
So, a better model can replace a generic
tool overnight, right? Maybe chat GPT6
comes out, it becomes amazing and you
worry about your customtuned special
model because it's out of juice. Or you
worry about your generic tool that was
just going to do one simple thing and
then Chad GPT just beats it. But if you
have vertical expertise backed by
vertical specific data, backed by your
vertical specific expertise encoded in
business rules, encoded in logic, that
takes years to build. that can't be
commoditized. That becomes an agent
framework that a better model
complements. You basically want to give
the organizational context to the model
as memory, as specialized tool calls, as
special workflows. And that becomes
something that a better model only
enhances because you have vertical
defensibility encoded in the context of
the orchestration layer around the
model. You don't need to worry about a
better model un undoing the value that
you've built. It actually just enhances
it because it's smarter, which enables
you to do higher quality work, tackle
more difficult problems, hit higher
completion rates. You get the idea. The
principle is to specialize and pick
vertical specific workflows. And if
you're building for the customer, if
you're in a B2B space, watch out for
those generic horizontal capabilities.
Those are very much at risk of
disruption right now. Generic tools do
get optimized out of existence because a
general model comes along and they just
optimize too fast and they were like
we're going to be a generic tool for
everybody for this specific thing and
like people don't care. As an example,
the chat with your PDF tools that's
going away because you can just upload a
PDF and chat with it now. See generic
tool. It solved a problem. It's no good
now. But if you have a vertical specific
workflow around health care or around
finance, that is not going away when a
better model comes out. And so this is
really good news for businesses that
have deep vertical expertise and
comparative advantage in specific
things. If you have comparative
advantage around a fintech thing or a
healthc care expertise, if you deeply
understand the regulatory environment
around real estate acquisition in
particular state, whatever your
expertise is, build vertical specific
orchestration frameworks, embed the
domain expertise inside the workflow as
much as you can, but keep it out of the
model because then you can swap the
model in and the model will learn your
verticals, compliance, terminology,
decision patterns, etc. uh from the
tools, from the context, from the
orchestration layer and be able to
operate. And so for example, a legal
compliance framework might understand
privilege. A healthcare memory system
might understand consent in a certain
way. Manufacturing agents can understand
equipment history. You get the idea. But
critically, these require the ability to
call in that context, not to assume the
model has it built in because you don't
have to assume that. you can get a
better model swapped in tomorrow and the
vertical specific piece can live inside
your business as part of that agent
orchestration layer. Principle number
five, compliance is a competitive moat
with agents. So we talk about the EU AI
act a lot because it is hugely relevant
for any business that that does any kind
of volume in the EU. Enforcement is
going to begin in the next year. You
have to have a compliance framework that
actually works. But if you do, it
becomes a moat. And I would assume that
you want compliance frameworks that
scale for you and anticipate emerging
regulatory environments in the United
States as well, which are happening on a
state-by-state basis and are very
complex. You want to think about the
regulatory infrastructure that enables
you to show auditability, traceability,
security first vendor integrations,
policy controls. Those are all things,
spoiler alert, that come from a solid
orchestration layer. They are not things
that the next model Chad GPT6 is going
to give you for free. They are things
that you have to build for your vertical
that essentially are like an extension
of your expertise into the regulatory
space touched by AI. You are building,
you are forging an orchestration layer
for agents that captures the regulatory
nuance of your world. I talked about
healthcare earlier. Well, HIPPA is
relevant in the US, right? You have to
think about privacy. How do you show
that your agentic framework is HIPPA
compliant? How can you show that at the
database level? How can you show that in
the run traces you have for agents? How
can you show that in the ability to
generate audit frameworks that people
who need to look at your systems and
sign off on them can easily verify and
understand. We are pioneering here,
right? We have not had a year of agent
maturity yet. We do now. Like the clock
started a couple of months ago and I am
issuing this warning because I have
heard too many times from leaders that
they are still picking the model. They
are waiting for agents to be fully
ready. Please don't wait. Please don't
wait. Principle number six, velocity
matters more than perfection. And I feel
like that's a wonderful segue. Right? If
you can get to something in 6 weeks that
gives you an 85% completion on a
particular workflow task, that is going
to beat a six-month planning cycle. That
is going to be thinking about this for
the budgetary 2026 cycle. Please think
about how you are building speed into
your systems. And I'm saying that on
multiple dimensions, both in the first
run of the agent and how you get it
going in your hiring plans and how
you're hiring for engineering talent to
get you there. And also think about the
fact that agents themselves are
accelerators. A good agent orchestration
layer delivers an accelerated impact
across not just the workflows it touches
but everybody around that agent because
they're no longer dealing with that.
Let's go back to Walmart. Walmart is no
longer dealing with the context
switching costs, the team drag that
comes from manually dealing with all of
those bugs because the agents are doing
it. there is a speed up effect across
the team. And so if you can invest in
velocity now to get into an agent
deployment, you are going to get
velocity speed up dividends after it's
deployed. And so you should think about
where in your business you have the most
blast radius effect if you deploy an
early workflow for agents. What is
something that is a huge pain point for
teams where you can deploy an agent and
you can see a tremendous speed up not
just cuz that task gets done. Sure, we
all know that's going to happen, but
because other teams aren't touching
things that have been incredibly
painful. So here here is the critical
piece. I want you to go back to the
beginning with me. What we have here is
a decision cascade and I want to walk
through all six and show you how they go
together. You are picking an
architecture because a model agnostic
orchestration survives churn. Inside
that architecture, you are deploying
memory because you are starting
institutional learning to generate
unclosable gaps with your competition.
Then inside the architecture with the
memory, you're automating workflows.
You're closing loops that generate
recurring compounding returns because
you're actually finishing those
workflows completely. and you are
picking those workflows around
verticality. You are building expertise
that you cannot commoditize. That entire
system is going to be built inside a
regulatory competitive moat. You are
going to think about the regulatory
environment you're in and you're going
to recognize that if you can demonstrate
compliance proactively with agents, you
have a moat just as you have a moat with
memory and you have a moat with
verticality. Compliance is a specialized
skill set that agents can learn and you
can teach it to them for your business.
Last but not least, moving fast enables
you to generate compounding returns not
just from the initial agent deploy, but
because this entire system generates an
acceleration for the business. And
that's what Walmart found and that's
what you're going to find. And so my
challenge to you is do not wait. Your
competitors who are doing this are not
waiting to be perfectly ready. They are
figuring out very scrappily how to apply
these kinds of principles to build
production agents. There is no technical
gap here. You may have a talent gap you
need to close and you can go and get the
talent for it, but there is no technical
gap. Agents are ready for production.
There is no reason to wait and people
who are asking me whether it is time are
already late. Don't be that person. You
have the principles here. you understand
how to think in agentic terms. If you if
you listen to this video a couple times,
you understand more about agentic
orchestration than 90% of the seauitees
I talked to. This is how you need to
think to build aic systems that last and
are sturdy both this year and into next
year and beyond. These are these become
systems that have ROI and investment
value, not just flashy demos. So build
to last, build now, and don't ask me if
we're ready for AI agents because it's
already too late. We are ready. We are
done. We're deploying. They're at the
Fortune 100 level. We are off to the
races.