Learning Library

← Back to Library

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.

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
0:00leaders. If anyone in your org is 0:02saying, "We're not ready for AI agents," 0:04they're wrong. If you are asking, "Are 0:07we ready for AI agents? Are AI agents 0:09mature enough for us?" You're wrong. 0:11While you are figuring out how to 0:12perfect your AI readiness deck, Walmart 0:15has automated 95% of their bug fixes 0:18with 200 specialized agents. AI agents 0:21are here in production at Fortune 100 0:24companies right now. Here are the six 0:27principles that Walmart and others have 0:29used to build AI agents into production 0:32workflows. And you can absolutely go 0:35after that this quarter. You do not have 0:37to wait. This is not a 2026 0:39conversation. One of the things that I 0:41worry about is that companies are going 0:43to take AI agents too slowly like 0:45traditional software. Like, okay, maybe 0:48we'll be ready. We'll add it in 2026. 0:50No. If you see an AI agent use case, go 0:53after it and go after it aggressively. 0:55So, what are those six principles? We're 0:57going to dive right in. Principle number 0:59one is architecture first thinking. So, 1:02executives tend to ask me, Nate, which 1:04model should we bet on for AI agents? 1:06Wrong question, guys. The right question 1:09is which architecture delivers the 1:12workflows we're looking for? Model 1:14advantage lasts maybe a quarter at best. 1:17Architectural advantage is something 1:18that persists for years. So the 1:20principle is this. You want to build 1:22model agnostic orchestration that isn't 1:25dependent on particular models to 1:27execute particular workflows. In 1:29particular, you want to make sure that 1:32your agents are selected, evaluated, 1:35swapped, and combined individually. And 1:37you do that based on the architecture 1:40and what it demands, the workflows, 1:42right? This comes back to your 1:44orchestration layer being the 1:45competitive bet, not the models. The 1:47models are replaceable commodities. you 1:50can build a model orchestration layer 1:52which is exactly what Walmart built and 1:54that is going to get you much much 1:56farther. So, you want to be in a place 1:58where you can deploy central 2:00orchestrators that manage specialized 2:02agents like Walmart's YB model is what 2:05they use. WIBY, you can look it up. They 2:08are designed to tackle a range of tough 2:12tasks with appropriate tooling, right? 2:15You can go after that by separating out 2:17your workflows into tasks, delegating 2:20those appropriately, aggregating them 2:22up, having appropriate guardrails and 2:24governance. It's system design. You're 2:26designing a system in which models sit 2:30and do particular tasks with particular 2:33tools. That way you will have 2:35infrastructure that enables you to 2:37switch models regardless of what comes 2:39out tomorrow. Microsoft cares so much 2:42about this they retired their autogen 2:43framework entirely to unify around an 2:46agent framework. 80% of enterprises out 2:50there are starting to get into 2:52orchestration layers roughly speaking. 2:54If you were going directly to models, 2:57you're building AI agents on sand. It's 2:59not going to last. Principle number two, 3:02learning compounds. And with AI agents, 3:05you need to invest early in order to get 3:08ahead of the market on compounded 3:11learning in 18 to 24 months. JP Morgan 3:14has had 18 months of institutional 3:17learning and you cannot buy, steal or or 3:20replicate that with better tech. They 3:22have spent 18 months on AI agents 3:24already in late 2025. Every day you 3:28wait, that gap widens. Right now, you 3:30may not be competing with JP Morgan, so 3:32you don't care. But that's what we're 3:33talking about when we talk about 3:35compounding learning. Memory systems 3:38need to be turned on early to turn 3:41adoption into a permanent advantage for 3:44you. Organizational learning 3:45accumulates. competitors are going to 3:47have to face unclosable gaps because you 3:51are going to have agents trained on more 3:53data than they have always. Always. So 3:56you want to be in a place where you can 3:58deploy early and capture significant 4:01percentage gains in accuracy in 4:04completion etc. You basically preserve 4:07an accuracy and quality gain as an edge 4:11over time because you are always getting 4:14the best models against a growing body 4:17of data that other people don't have. 4:18This gives you asymmetric strategic 4:20value. So you want to implement memory 4:23augmented agents inside an orchestration 4:26layer that learn your organizational 4:28context that track the information that 4:31matters when it matters and how it 4:32relates to decisions. And yes, the 4:34examples I'm giving you, that's exactly 4:36what's happening in task appropriate 4:38ways. In this case, for Walmart, it was 4:40triaging bug tickets. That was really 4:42important. That's what they focused on. 4:43Focus on systems that are able to absorb 4:46and understand meaningful workflows. 4:49Make sure they have clear terminology 4:51they can learn from, like dictionaries. 4:52Make sure they have clear workflow 4:54patterns. to make sure they have really 4:56really crisp precedents to work against 4:59how things have worked in the past, how 5:00you've done b business logic in the 5:02past, how you handle data ambiguities, 5:05and make sure they have exception paths. 5:07200,000 JP Morgan employees have gotten 5:11450 use cases into AI agent land. 5:17Basically, they've they've uploaded 5:18those use cases. They've had AI agents 5:20start to get to get into those use 5:21cases. A system deployed today will 5:25understand your organization as well as 5:27JP Morgan does now for those 450 use 5:29cases in 2027. Right? Don't wait. That's 5:34the whole point of this video. Don't 5:36wait. AI agents are ready now. That is 5:38one of the big things that has shifted 5:40in the past few months is that there is 5:42no reason to wait. It isn't even as 5:44expensive as it used to be. AI agent 5:47SDKs are out there for major model 5:49makers all over the place. They're out 5:50there for cloud providers. You can 5:52assemble AI agent frameworks. Your team 5:55can. Principle number three. Let's talk 5:57about workflows. I've been referencing 5:58it a little bit. Demos are going to get 6:00you applause or get your engineers 6:02applaud, but but the workflows that I'm 6:05talking about are really aimed at real 6:07ROI. And that is the difference that 6:09most executives talk about when they get 6:11AI right. Enterprise AI spending is 6:14scaling by like triple digits a year in 6:16percentage point terms. But it's scaling 6:19because boring automation prints return 6:22on investment. It really like the the 6:24return on investment is skyhigh if you 6:26can pick specific automations for 6:29defined workflows and stick agents 6:31against them. And so you should be 6:33looking not at feature counts, not at 6:36the number of agents you launch, not at 6:38login metrics, not at the number of 6:41tickets even. You should be looking at 6:43the rate of correct completion for 6:47workflows that matter to the business. 6:49Can you get to 90% plus AI automated 6:53completion for workflows that matter to 6:55the business? Recurring automation will 6:58compound in value in a way that flashy 7:00demos will never do. If you can get to, 7:04you know, call it 30% automation in 7:06month one and you have a dedicated team 7:08working to drive that number, the value 7:10just compounds. is that team bites off 7:12more edge cases and is able to knock out 7:14that workflow and they go on to the next 7:16workflow and eventually it gets easier 7:17and easier as you start to stamp these 7:19out. Identify the highfrequency 7:21high-cost workflows first. Of course, 7:23you want to focus on workflows that have 7:26enough defined inputs that you can have 7:29tools called against defined inputs and 7:32the agent can make correct decisions at 7:34a very high rate to fully complete the 7:37workflow. If you can't get that, pick a 7:39different workflow initially because if 7:41you pick a high ambiguity workflow 7:43initially, it's going to be hard to get 7:45success. You'll grow discouraged and you 7:47probably won't go after workflows with 7:49the aggression that you need to. 7:51Principle number four, vertical 7:53defensibility. 7:54So, a better model can replace a generic 7:57tool overnight, right? Maybe chat GPT6 8:00comes out, it becomes amazing and you 8:02worry about your customtuned special 8:04model because it's out of juice. Or you 8:07worry about your generic tool that was 8:10just going to do one simple thing and 8:13then Chad GPT just beats it. But if you 8:16have vertical expertise backed by 8:18vertical specific data, backed by your 8:20vertical specific expertise encoded in 8:23business rules, encoded in logic, that 8:26takes years to build. that can't be 8:28commoditized. That becomes an agent 8:31framework that a better model 8:34complements. You basically want to give 8:36the organizational context to the model 8:39as memory, as specialized tool calls, as 8:42special workflows. And that becomes 8:45something that a better model only 8:46enhances because you have vertical 8:49defensibility encoded in the context of 8:52the orchestration layer around the 8:54model. You don't need to worry about a 8:57better model un undoing the value that 9:00you've built. It actually just enhances 9:02it because it's smarter, which enables 9:04you to do higher quality work, tackle 9:06more difficult problems, hit higher 9:08completion rates. You get the idea. The 9:10principle is to specialize and pick 9:13vertical specific workflows. And if 9:15you're building for the customer, if 9:16you're in a B2B space, watch out for 9:19those generic horizontal capabilities. 9:21Those are very much at risk of 9:22disruption right now. Generic tools do 9:25get optimized out of existence because a 9:28general model comes along and they just 9:29optimize too fast and they were like 9:31we're going to be a generic tool for 9:32everybody for this specific thing and 9:34like people don't care. As an example, 9:36the chat with your PDF tools that's 9:38going away because you can just upload a 9:40PDF and chat with it now. See generic 9:43tool. It solved a problem. It's no good 9:45now. But if you have a vertical specific 9:47workflow around health care or around 9:49finance, that is not going away when a 9:52better model comes out. And so this is 9:54really good news for businesses that 9:56have deep vertical expertise and 9:58comparative advantage in specific 10:00things. If you have comparative 10:01advantage around a fintech thing or a 10:03healthc care expertise, if you deeply 10:06understand the regulatory environment 10:07around real estate acquisition in 10:09particular state, whatever your 10:10expertise is, build vertical specific 10:13orchestration frameworks, embed the 10:15domain expertise inside the workflow as 10:18much as you can, but keep it out of the 10:20model because then you can swap the 10:22model in and the model will learn your 10:24verticals, compliance, terminology, 10:25decision patterns, etc. uh from the 10:27tools, from the context, from the 10:28orchestration layer and be able to 10:30operate. And so for example, a legal 10:32compliance framework might understand 10:34privilege. A healthcare memory system 10:36might understand consent in a certain 10:37way. Manufacturing agents can understand 10:39equipment history. You get the idea. But 10:42critically, these require the ability to 10:44call in that context, not to assume the 10:48model has it built in because you don't 10:50have to assume that. you can get a 10:51better model swapped in tomorrow and the 10:53vertical specific piece can live inside 10:55your business as part of that agent 10:57orchestration layer. Principle number 10:59five, compliance is a competitive moat 11:02with agents. So we talk about the EU AI 11:06act a lot because it is hugely relevant 11:08for any business that that does any kind 11:11of volume in the EU. Enforcement is 11:13going to begin in the next year. You 11:15have to have a compliance framework that 11:16actually works. But if you do, it 11:18becomes a moat. And I would assume that 11:20you want compliance frameworks that 11:23scale for you and anticipate emerging 11:25regulatory environments in the United 11:27States as well, which are happening on a 11:29state-by-state basis and are very 11:30complex. You want to think about the 11:33regulatory infrastructure that enables 11:36you to show auditability, traceability, 11:39security first vendor integrations, 11:41policy controls. Those are all things, 11:44spoiler alert, that come from a solid 11:46orchestration layer. They are not things 11:48that the next model Chad GPT6 is going 11:51to give you for free. They are things 11:54that you have to build for your vertical 11:56that essentially are like an extension 11:59of your expertise into the regulatory 12:02space touched by AI. You are building, 12:04you are forging an orchestration layer 12:07for agents that captures the regulatory 12:10nuance of your world. I talked about 12:11healthcare earlier. Well, HIPPA is 12:13relevant in the US, right? You have to 12:15think about privacy. How do you show 12:17that your agentic framework is HIPPA 12:19compliant? How can you show that at the 12:21database level? How can you show that in 12:23the run traces you have for agents? How 12:26can you show that in the ability to 12:29generate audit frameworks that people 12:31who need to look at your systems and 12:33sign off on them can easily verify and 12:36understand. We are pioneering here, 12:38right? We have not had a year of agent 12:40maturity yet. We do now. Like the clock 12:43started a couple of months ago and I am 12:46issuing this warning because I have 12:48heard too many times from leaders that 12:50they are still picking the model. They 12:51are waiting for agents to be fully 12:53ready. Please don't wait. Please don't 12:55wait. Principle number six, velocity 12:57matters more than perfection. And I feel 13:00like that's a wonderful segue. Right? If 13:02you can get to something in 6 weeks that 13:04gives you an 85% completion on a 13:07particular workflow task, that is going 13:09to beat a six-month planning cycle. That 13:11is going to be thinking about this for 13:13the budgetary 2026 cycle. Please think 13:17about how you are building speed into 13:19your systems. And I'm saying that on 13:20multiple dimensions, both in the first 13:22run of the agent and how you get it 13:24going in your hiring plans and how 13:26you're hiring for engineering talent to 13:28get you there. And also think about the 13:31fact that agents themselves are 13:32accelerators. A good agent orchestration 13:35layer delivers an accelerated impact 13:38across not just the workflows it touches 13:41but everybody around that agent because 13:44they're no longer dealing with that. 13:46Let's go back to Walmart. Walmart is no 13:48longer dealing with the context 13:51switching costs, the team drag that 13:54comes from manually dealing with all of 13:56those bugs because the agents are doing 13:58it. there is a speed up effect across 14:01the team. And so if you can invest in 14:03velocity now to get into an agent 14:07deployment, you are going to get 14:09velocity speed up dividends after it's 14:12deployed. And so you should think about 14:15where in your business you have the most 14:18blast radius effect if you deploy an 14:21early workflow for agents. What is 14:23something that is a huge pain point for 14:25teams where you can deploy an agent and 14:28you can see a tremendous speed up not 14:30just cuz that task gets done. Sure, we 14:32all know that's going to happen, but 14:34because other teams aren't touching 14:35things that have been incredibly 14:36painful. So here here is the critical 14:39piece. I want you to go back to the 14:41beginning with me. What we have here is 14:44a decision cascade and I want to walk 14:46through all six and show you how they go 14:48together. You are picking an 14:49architecture because a model agnostic 14:52orchestration survives churn. Inside 14:54that architecture, you are deploying 14:57memory because you are starting 14:58institutional learning to generate 15:00unclosable gaps with your competition. 15:03Then inside the architecture with the 15:05memory, you're automating workflows. 15:07You're closing loops that generate 15:09recurring compounding returns because 15:13you're actually finishing those 15:14workflows completely. and you are 15:16picking those workflows around 15:19verticality. You are building expertise 15:21that you cannot commoditize. That entire 15:24system is going to be built inside a 15:29regulatory competitive moat. You are 15:31going to think about the regulatory 15:33environment you're in and you're going 15:34to recognize that if you can demonstrate 15:37compliance proactively with agents, you 15:40have a moat just as you have a moat with 15:42memory and you have a moat with 15:44verticality. Compliance is a specialized 15:47skill set that agents can learn and you 15:49can teach it to them for your business. 15:51Last but not least, moving fast enables 15:55you to generate compounding returns not 15:57just from the initial agent deploy, but 16:00because this entire system generates an 16:03acceleration for the business. And 16:05that's what Walmart found and that's 16:07what you're going to find. And so my 16:09challenge to you is do not wait. Your 16:11competitors who are doing this are not 16:14waiting to be perfectly ready. They are 16:16figuring out very scrappily how to apply 16:20these kinds of principles to build 16:22production agents. There is no technical 16:25gap here. You may have a talent gap you 16:28need to close and you can go and get the 16:29talent for it, but there is no technical 16:32gap. Agents are ready for production. 16:34There is no reason to wait and people 16:37who are asking me whether it is time are 16:40already late. Don't be that person. You 16:43have the principles here. you understand 16:45how to think in agentic terms. If you if 16:47you listen to this video a couple times, 16:49you understand more about agentic 16:51orchestration than 90% of the seauitees 16:54I talked to. This is how you need to 16:57think to build aic systems that last and 17:00are sturdy both this year and into next 17:03year and beyond. These are these become 17:05systems that have ROI and investment 17:08value, not just flashy demos. So build 17:10to last, build now, and don't ask me if 17:15we're ready for AI agents because it's 17:18already too late. We are ready. We are 17:19done. We're deploying. They're at the 17:20Fortune 100 level. We are off to the 17:22races.