Future of AI Agents – Andrew Ng & Harrison Chase

Speaker(s): Andrew Ng (Founder, DeepLearning.AI); Harrison Chase (Co-Founder & CEO, LangChain)
Session: Interrupt 2026 · Day 2 (May 14) · ~1:00 PM PT
Source: in-person audio recording, transcribed locally with Whisper large-v3.

Summary

In a fireside chat with Harrison Chase, Andrew Ng (Founder, DeepLearning.AI) reflects on the rapid evolution of coding agents and how he now builds with small, high-context generalist tasks given guardrails and freedom to run. He argues that people from any background (engineers, product managers, marketers, operations) can learn to build, and that the real leverage comes from a rich, growing set of composable building blocks (APIs, SDKs, frameworks) rather than from any single tool. On enterprise adoption, he warns that running thousands of AI experiments often yields incremental efficiency rather than transformation; the real wins require top-down workflow redesign (his loan-underwriting example) and chasing growth (20-50%) over cost savings. He emphasizes preserving optionality given rapidly changing leading models, the value of vendor-neutral tools and open-weight models, and a coming wave of unstructured-data re-architecture work, while noting permissions were designed for humans, not agents.

Key Points

• Coding agents have evolved faster than expected and change rapidly; Ng favors setting up many small, high-context, highly generalist tasks with standard guardrails that can 'run like crazy'
• People from any background can build with AI; he sees managers, product managers, marketers, and operations people learning to code, not just engineers
• Leverage comes from composing a rich and growing set of building blocks (APIs, SDKs, frameworks); the value grows roughly exponentially with the number of available blocks, like Lego pieces of many shapes
• A challenge is that code agents often don't know how to use newer building blocks (knowledge-cutoff issues); documentation and a 'Compact Hub'-style approach help agents call newer APIs accurately
• Enterprise lesson: many AI experiments deliver incremental efficiency, not transformation; real value requires someone with broad scope to redesign the entire workflow top-down (loan underwriting reimagined as a 10-minute 'get-approved' product)
• Pursue growth over cost savings: cost savings have a ceiling, but growth has almost no practical ceiling; treat ideas as a portfolio rather than one wild swing for the fences
• Given rapidly changing leading models, optionality is very valuable: he favors vendor-neutral tools and open-weight models and personally avoids long lock-in contracts even when offered 20-30% discounts
• Education is being transformed (Coding.AI: a conversation/simulated video call rather than a course; clicking into JavaScript-driven videos is live today), and unstructured data (text, images, PDFs, audio, video) re-architecture is a coming large-scale challenge, with permissions designed for humans not agents

Notable Quotes

It's not a course, it's a conversation

because it could only save so much money, but growth has almost no practical ceiling

in moments of uncertainty like this, optionality is very valuable

the permissions were designed for humans, not for agents

Full Transcript

Show the full timestamped transcript (auto-generated; lightly cleaned)

[00:00] Welcome back everybody. We have an absolutely stacked back half of intro up, so I'm really, really
excited to get things underway in this second half of the day. Next up is Andrew Ng, founder of Deep
Learning, who will be having a backside show perhaps. Andrew has been thinking about the future of
software engineering longer than most, and lately he's been writing about something that comes right
to the heart of what we're all building towards. While leading a thousand AI experiments, Bloom
hasn't actually translated into transforming business impact, and what workflow redesign actually
looks like when you get it right.

[00:41] Please welcome Andrew Harrison. Thank you for being here, Andrew. It's great to have you back. We're
here too.

[01:13] It's good to be back. It's always such a cool gap when we get together. I was telling you this
earlier, but Andrew's fireside chat last year was the most liked and the most watched on YouTube
afterwards, so we're thrilled to have him back for us here too. Woo! Yeah. Thank you. Maybe jumping
off of that, in the years since we've been here, I feel like there just been a ton that's happened
in the AI space. What has happened faster than you would've expected, and what's been slower?

[01:44] So I think that, um, that part has exceeded my expectations. And then also, the do-it-yourself
analysis is also often faster than that, including the job conference. So thank you. And on the one
hand, coding agents are a question of . And the concept of coding agents, I know that is that
everything changes every three months or whatever, and that type of thing is probably true.

[02:15] But that does be true of the coding agents. It feels like the concept of coding agents and in
certain cases, I think about six months ago, I was almost all . These days I'm still all the . But
it makes the CLI and . So it makes the coding agents change rapidly. I wouldn't have guessed a year
ago that .

[02:46] So this . Like when you're in the office, all these . I find myself setting up very small tasks,

[04:25] things that have grown one to 10 million years in the team, but often generalist, high context,
highly generalist, that are given the standard 31 guardrails, but in which they can just run like
crazy and go and shit code, and even drive decisions like writing marketing cards, like traditional
on-site information. By definition, let's say you have a team that needs software gaming, product
management, you build up the terms of service,

[04:56] you need some coffee, you need some design, say you need five functions for a detective, you need a
team of two humans, then by definition, by the principle, these two humans have to play more of the
same kind of role as a human. What you could do is, it turns out, I don't think I'm very good at
this stuff, but when I use AI, I'm still good at it. Right? So, there are some good marks that are
not as bad as the ones I have. So, I find these small things of high context generally is that it's
really technical,

[05:27] but able to use fancy technology. There are two million and a half million dollars of software. Then
you have engineers, you know, and these AI needs to take the first draft of the terms of service,
and then take the role of the lawyer to find the big policy and put it to the top. So, I find that
these processes allow for a lot of things. That's the first thing I would say. And what's the race
background for these folks? Are these engineers by background? Are they coming in from different
disciplines and learning to code for the first time? What are you seeing in this? Yeah, so, a lot of
the people I work with are both in the junior technical facilities.

[05:59] And then also, when you think about general disciplines, the second of those are business skills,
often, they're not supported where they're doing the training. I think you can substitute them in
any direction. I've seen people, a lot of managers, they come much further in coding than they are
in the top three. I think there's so much of the idea of coding, engineers need the natural
advantage, understanding the frontier tech. So, I see engineers' failures as well as success as
people's already.

[06:31] But there's definitely a small amount of the product managers that see, you know, marketers, or
whatever it's called, in effective ways as operations people, and some people are more of a product.
I think that's also for people, you know, any background, they're going to do a lot of things. But,
you know, not just something like a body of things, but that's not exclusively people who are going
to do the engineering right now. But we encourage everyone to learn the background to see the future
of these things. What advice would you have for folks who are trying to get more into this, in this
new software engineering space,

[07:02] whether it's particular tools to try, or mindsets to have, or skills to pick up? Yeah, so, there's a
lot that I've been thinking about the future of software engineering. This is a mental problem. And
I think that's a lot that I have to deal with. Which is that, because of a lot of providers, of many
tools, like RAD, and other frameworks, things like the e-mouse, Godrail, it turns out that,

[07:33] oh, that's right, so these are paid-back-worthy products. They're also non-paid-back-worthy
products. Things like user-based control, or identity-off mechanism, you know, turns them back into
assistant data, assistant software, and so on. So, I think that the computer space will always have
a wonderful set of building blocks. And when it comes to coding, building blocks are proliferating.
There's more and more people who are building for the source, or the time-sharing API, or whatever
the best way to call it. So, there's this wonderful building block

[08:03] all around us. And so, I find that developers that have a good box safety of enough building blocks,
they're not going to put them together in complicated, and many of these are really back-to-the-
work. And maybe not necessarily with Lego bricks, if all I have is like a white-colored Lego brick,
I could build some stuff that's not that interesting. But I think some black ones, yellow ones,
brown, green ones, some squid-like pieces of Lego, then the furniture built

[08:33] grows commonly or grows exponentially as a function of the number of blocks that are built by now.
And I think that a lot of building blocks do not have access to these things back here. I find that
the developers have a good sense of what this building blocks. And even though I talk about the form
of the building blocks, the poses, the people, the master, this wonderful building blocks is
provided by a constant flow. And then there are challenges in these code agents that are actually
the central building blocks of software.

[09:05] One challenge that code agents have is a lot of building blocks that are so huge that the code
agents do not know how to use them. Until recently, the leading code agents were nanomathemathems,
which is often knowledge-cut updates and a lot of new tools. So you just have no idea how to call a
nanomathemathemathem, even if you have a nanomathemathemathem. So one project that one of my friends
wrote for us that I'm passionate about is a project called, which is kind of a stack of the flow of
AI agents.

[09:37] What AI agents generate is documentation. What all of the latest APIs, SDKs, building blocks can use
as a way for agents to give feedback to the documentation. And so it turns out, the number of APIs
that I personally use, that I find the syntax struggling, that I find that by using Compact Hub, the
laser starts, and then the code agent accurately makes all of these APIs possible.

[10:07] So actually, let's call it my own. So I'm going to go ahead and start. Maybe if you notice there, we
also launched something called Compact Hub. So we're colliding on names, but on very different types
of contacts. And so, yeah, it's much more useful for working with code agents. They should take
advantage of that. And then deep learning. So I think we at LinkedIn were the second people to hide
open AI, obviously, to do a deep learning class. And I know I've talked to so many people who have
heard of LinkedIn through deep learning,

[10:38] and I'm sure a lot of the audience members here as well. So if you haven't read that deep learning,
you could absolutely go take some courses. Maybe on that note, how do you think about education
changing in this new world of AI? Have you started making any of those practices into how the deep
learning courses are run, or how do you think about that? Yeah, so we're kind of one of the greatest
in the education space. So certainly, the thing that's clear is that what people have to learn has a
case in mind.

[11:09] I think the developers, who are coding in this, learn these building blocks, maybe learn some
product management, and some general statistics, what deep learning is changing. And even the
radicals there are, I think, are the best of all the managers. But separately from what the learning
is, the delivery of the learning is. And it feels like we've been thinking about how we're learning
on school for a long time, and it feels like there's actually talk about deep learning.

[11:39] One thing we launched just several weeks ago is a new website in Google called Coding.AI, in which
the vision was, rather than taking an online course, come and have a conversation. It's not a
course, it's a conversation, where the experience we tried to build was for you to come and get on a
simulated video call, with a link, say, where if you do that, link, and then you say,

[12:09] I'm listening to you talk and present to you one-on-one video similar to a video call about a topic
that I'm creating, so I'm going to do that, and show you something interesting. Or if you want to
interrupt, I can go and meet you, or interrupt by email, and ask a question, and I'll get you to do
so. What I can actually, firstly, bring out a lot here is replacing videos and slides with
JavaScript. And what that means is, instead of a video, when I present, when I demo something in a
video, then you can click into the video

[12:41] and type your own props, or type your own queries into the video, where you can have a set of static
video displays, the video then will design the discase always if you do that. So check the coaching
that we have at the Wednesday, that they kind of have a conversation with me over my AR, without
presenting to you the AR form, how to use comments on that. I think we're still, we're actually
trying to be so thin with these views. Is that clicking into the video and taking,

[13:11] is that live there, or is that more of a future direction? That's live. So imagine instead of me
sharing the video, I am, you know, JavaScript sharing the video, and so you can, it's a matter of
course, you share, this is running JavaScript on a . Yeah. What's the question? Any of you, I love
your feedback, but it is like, I think the transformation of, you know, education has been over,
right? I think something has come,

[13:41] like today, taking online courses, you know, I wish we had something way better than online courses,
and I would say that something better than the courses we've had at Cisco, at the Washington
Tractor, it turns out a lot of our courses at this moment, we just want to be able to transform and
how about we share the truth with the public, and I find that rather than just having videos, which
is mostly what we had a decade ago, we now build much more

[14:11] into practice and visualization, so one more fun side of television that causes them to take a go,
but I think the transformation, I actually spend a lot of time making a lot of them. Maybe going
from software engineering to everywhere else, how do you see enterprises adopting AI, and is it
faster than you expected, slower, what's the right way to do it, what lessons can people learn? So I
think every enterprise, I guess all of yours,

[14:41] makes sense of AI adoption. One thing we've seen is, one of the many businesses, one of my teams, AI
Aspire, which is an AI advisory firm, we have a lot of companies, this is my company, Chris is my
company, we talk to launch businesses all the time, so we have a lot of people that are working with
the Fortune 500, GTC, and all the best strategies in the construction industry. So sometimes there's
some things that are interesting, but sometimes there's some things that are not. So I think we
should keep on investing in

[15:11] the Fortune 500 innovation that's a thousand dollars in the future. And for the most part, it's not
paid for. So CEOs and boards are asking, where is the R&R going? I think we should keep on investing
in Fortune 500 innovation. Let's keep on doing that. But it turns out that Fortune 500 innovation
often results in fun solutions that drive incremental efficiency for the enterprise, which is
actually a good thing, but not the transformation, not the global transformation that we are testing
on the market. So we also must keep on protecting the future of Fortune 500. So let's illustrate

[15:41] this with an example. My team is working with a number of banks. We actually do a lot of financial
services. But we're working on the banks, and when you think about the process of underfunding a
loan, there may be five steps. There are marketable loans, product, IT application, we're here to
prove the loan, we define the religion, so basically that's sort of the process. Number of things
that we'll see is that the step in the middle of building approval, you could use a lot of money to
do that. And if the ultimately that

[16:11] then instead of a human spending an hour doing the loan application, the AI could do it, and that's
great, we should actually do that. But it turns out that the entire process of underfunding a loan
stays the same, except for ultimately what was previously one hour of human time. That's a small
incremental difference. So what a number of banks have said is, you know what, instead of doing this
efficiency game, which is what it's all about, let's rethink the entire workflow, and market a get-
approved

[16:41] and 10-minute go-to-go. Because rather than waiting around for 30, but humans who need an hour, we
can send the loan application to AI right away for a decision. But the challenge of implementing
this in one of the businesses is, is that someone with a robust scope to rethink and redesign the
entire workflow, because now there's a market, a get-approved and 10-minute product, you have to
route the application to approval, not within a day, but right away.

[17:11] So not basically from this Then yes, AI can make the initial decision, and then final execution as
well. And so I find that innovation is really valuable, generates lots of ideas, but often it takes
that as the top-down notion of having someone with a robust scope to change how all of these steps
are implemented. And I find that many businesses talk about

[17:41] cost savings, you know, that's not what I'm doing, but the more I try to push for more imaginative
things to do with AI, which is driving growth, because it could only save so much money, but growth
has almost no practical ceiling. So the more exciting ideas I find, usually the way to drive
business growth is to have a

[18:11] business that's really innovative, that's really innovative and that's really exciting to have in
the future. So I think that's the way to drive that business growth. But, when you can automate
customer service,

[18:41] or automate an off-hand hardware, then the ability to serve customers many more much faster, that
delivers a more intellectual customer experience in a drive-through. I think that actually speaking
with a number of businesses that are working on automating the drive-through order process, I think
that also results in a more innovative way of doing

[19:11] that. I'm seeing that there are more and more of these examples There's some I know about that I
don't have permission to talk about, but I'm actually pretty confident with the data that I've got.
I think that's a good

[19:41] way to think about the process of doing that. I think that's a good way to think about that. I think
that's a good way to think about this. I think that's a good way to think about

[20:11] the process of doing this. I think that's a good way to think about this. I think that's a good way
to think about Oh, yeah. There's actually one thing out there at Yale. Sometimes driving incremental
gains is all there,

[20:41] and driving transformative gains. Because if you tell someone to improve their business results by
2% in Yale, then they're like, all right, my boss is telling me to work 2% harder, or 5% harder, or
whatever. But if you try to search the ways to drive 20% business growth, or 50% business growth,
then you can't just get everyone to work 50% and be confident in trying to . That confidence. Just
one lesson of it. At AI Aspire, we've had many businesses literally send us spreadsheets

[21:14] with hundreds of ideas. One financial institution sent us a spreadsheet with over 300 ideas and
processed the whole of these 300 ideas, which makes the quick little capital to invest in. And it's
just so that the analysis is really difficult. I wish I was smart enough. I could say, oh, this
idea, this idea. But I find that when faced with exactly like this, often based on top-down and
bottom-up emotion, there's actually a lot of work to do the technical analysis

[21:44] to figure out what is possible. And also the business analysis to figure out which ones could drive
the inflation and where not. There's actually a lot of work, the technical and business analysis,
but at the bottom, there was a small handful of tests that could very easily . And these swings for
the fences take things. You're often seeing these being the top-down ones. That's . Yeah. I find
that businesses hopefully not take one wild swing for the fences. There's more of a portfolio of a
handful of comfortable beds

[22:16] where if anyone pays off, it will be equal for the business. But it turns out one thing I love about
Virginia College, a bunch of times I've been there, full-time, all-time, so the cost of full-time is
a lot of sales. You can't do everything by a $100,000 budget and at some point putting the resources
behind every one of the small portfolios that have those branches. So because of the resource
allocation needed at that level,

[22:47] it often takes a little bit more of a top-down . One of the things that I feel like has been talked
a lot recently about enterprise-bound with AI is forward-deployed engineering. Will every company
have forward-deployed engineers? Why do you think they're so impactful and how do you see this
playing out? So I think the Silicon Valley does definitely have a role in the FDEs. I know Aaron was
in the stage.

[23:18] He was a very thoughtful team member. So I think FDEs are a great option for many businesses. But I
think looking to the future, what do you think is the ratio of, I think, FDEs in the company versus
the number of just AI engineers that are going to come? I think that most businesses will have a lot
more in-house engineers and a smaller team of FDEs maybe. So that's why I like FDEs. I'm excited
about the growth. That's how more people get jobs is FDEs.

[23:48] But I think the hype is also as strong because of the great success. But it is a good thing. But
don't be getting workflows. It's hard. It's hard to understand your business. It requires customer-
facing skills, often to make a viable . Observability, how you guys work with customers, push back
if something's not actually technically feasible, work with the stakeholders, ultimately helping to
change management.

[24:19] So this is a very valuable role that takes technical judgment and technical use of data and really a
stable . There's one other thing I see, which is challenging for a lot of businesses. Is there any
way to get the end of the curve? That sounds challenging, depending on what the industry is trying
to do. Because what we see in AI is that the leading AI model rapidly changes. So I have no idea
what will be the leading AI model here from now.

[24:54] I'm actually not at all sure what will be the leading coding agent here from now. And so in moments
of uncertainty like this, optionality is very valuable. So candidly, many vendors are coming all
about this and are offering 20, 30% response to start every year. I'm not giving any advice, just
saying what I do. I personally almost never sign wrong with anyone. In fact, the novice is . Because
I value that optionality.

[25:25] So whatever they want will be the best . And then when you work with FTEs, one question that I think
this is asking is, when you have a handful of FTEs for one company in your company, how much does
that mean? And then everything with one AI model, whatever, what's your optionality? And this is why
I put a couple of thoughts

[25:57] into this. I think I personally have used a bunch of . I think has done a great job. It is so easy
to use. I think those types of vendor-neutral tools are very valuable for preserving the same
optionality. And the vendors are great. Work with vendors. But preserve optionality for yourself and
the model. On the topic of vendor-neutral, especially in the model space, one of the things we
talked about a little bit throughout the past few days

[26:28] is kind of like open source models. How do you see those progressing and see them, how do you see
them relative to the FTE model? Yeah, it's been fascinating how well it's remained persistently It's
behind the FTE models. The FTE models are expensive enough that for many of these cases, my teams
use broadly open-weight models. Sometimes my teams don't. So I hope that it all people support the
open-weight model. This is a great question.

[27:56] The ASI, we expect to see a large business, very hard hit from this, rethinking the data object.
Because over the last 10-20 years, we've put so much effort into organizing our structured data,
tables, information data, spreadsheets, that's very important. But now that they are, it's process
and structured data, text, images, PDF files, audio, print video, organizing that, getting to the AI
for reasons like time, but first,

[28:26] but the value is certainly much more valuable than it used to be before. And frankly, I've done a
lot of work in the market. There are many vendors starting to talk about dealing with unstructured
data, but I've not been assigned a single good solution that I'm actually really satisfied with. So,
within my experience, I've been inspired, with the brand, such a crazy experiment, you know,
building, of,

[28:57] re-architects of a real place. In other words, I've been talking about this, I've actually spent a
lot of time thinking about how the re-architect problem, or the unstructured system, begins in
agents, but in agencies in general. I actually foresee that, just as many businesses have very large
data architecture problems, or data architecture, kind of, work, that people have structured data,
over the next few years, there will be very large, you know, let's say tens of millions,

[29:27] hundreds of millions of projects, in many businesses, I think that they are, you know, significant
ways of all day, by, you know, what age. What's the, what's the issue with their distinct data
architecture that doesn't make it pay-out by your own money? Boy, I know. Transplantation,
governance, pages on the place, notices, you know, some of this, you know, some of that, the
permissions were designed for humans, not for agents, so agents would carry my permissions. How do
you manage governance?

[29:58] I'm observing this. I feel like, you know, we've all seen, right, so many businesses that have
massive, you know, buckets of tons of PDF files, and you don't have to look at the process, because
you see in financial services, a lot of documents are contained for compliance purposes. So,
previously, there was no point looking at it, there was no time, but sorting that out, and I have to
go back to the site, and do that again. Oh, by the way, one small thing, this doesn't actually have
full data on question, but I think CJS can do this. Yeah. I would say,

[30:28] one little lesson I've learned, for AI, I think, focus is actually important. I, I first use one
with email, but, you can see this, because, I, yeah, we all love relational iterating, right? But, I
find that, when I'm iterating a program, I'm iterating it properly, the need to redesign the
database, right? It's so important. And, we've all had that, you know, one or a hundred times,
there's a constant need to, to do a database migration, and do something similar, like,

[30:58] create a separate database, right? Almost never happens. But, the fact that it, almost never
happens, but doesn't, never happens, is, you know, a little annoying. So, I find that, having an IOS
sequel, right, and, some other data, I would want to, you know, build a database, and then, you
know, figure this out, and then, rather than a database, it doesn't really work that much. NoSQL
doesn't always scale with a modest production. It does, very large production,

[31:28] but it does, as far as they, eventually, you know, at least, all the initial releases, must be, very
scalable solutions. But, I think, NoSQL, is more scalable, than most people realize. It drives that,
basically, maybe, I'll just get, so frustrated, if I design, you know, some database, and then, oh
shoot, I want to add a field. It's just, right, so annoying, that it de-facto, is terrible.

[31:58] So that, so I find that these are examples of the workflow, that we'll all need, to drive fast
iterations, to combat the fact that the IP agents, decode the device. So, let's not get slow. Yeah,
I think you're talking about getting agents to change, not only our, what we do, but also the
technology forces that are good for, for what the build on top of. I think I speak for everyone,
when I say, thank you so much for being here, thank you for sharing all your thoughts, and thank you
for all you do for the ecosystem. Thank you.