Hey.

Welcome back in here.

Another awesome talk.

>> Yeah, that was awesome.

>> I mean, it's really cool to see just

kind of how everyone's approaches

differ. Like everyone's not so your

approaches differ, but like their

opinions on where things are really at

and how to uh

implement

always seems to be like a little bit

different.

>> Yeah, yeah. I mean, generally we had two

incredible days. We had 14 sessions. We

had Uber here, Stripe, Cloudflare, just

saw Monzo, Harvie, AWS folks, Genentech.

So,

a crazy lineup. And I'm also very

appreciative that after 3 hours so many

of you are still sticking around. So,

thanks for tuning in from all those

different platforms. Um

And a quick note again. So, we will

follow up with the recordings. If you

joined via like Extra so, make sure to

go to background-agents.com/summit.

Um register with your email so that we

can actually share the recordings with

you. We will also share this as a

Spotify playlist so that you can listen

to it on the go cuz YouTube requires you

to be premium if you don't watch it. Um

and we'll probably also distill this

into a notebook LM um podcast all those

14 talks into one

um synthesized message basically.

So, um let's close the loop. Actually,

nice really nice framing to that. Um so,

we opened the summit with three pillars.

Um infrastructure was one, security was

the other, and then leadership and like

how to roll this out um the last one. Uh

and we also spoke about the shifts in

the SDLC and the how the role of the

human changes in the in the keynote. So,

let's actually close the loop, uh talk

about what we learned, synthesize some

of the learnings, and um

how and who demonstrated what in

practice. So, let's kick it off. Well,

curious what you what you think.

>> Yeah, so I thought it was really

interesting to hear how environment,

governance, and security kind of all

coalesce, right? Um

I mean, across talks from yesterday and

today, like dev boxes being so

important,

um

and the security that dev boxes kind of

gives you so long as they're within your

own cloud,

uh really seemed so important.

Um

and going from that to

implementing something similar within

AWS and you utilizing all these crazy

divergent AWS tools to build something

similar.

Um

basically, it's it's

it's really interesting to see how

the architecture that is the

architecture that's emerging

is

relatively consistent, but the

insights and the tweaks along the way

are all so so different. Um

Do you want to add a little more color

to that, Phil?

>> I think you're spot-on. I I would

totally agree with that.

And it's funny that like

the two pillars that we had in the

beginning, I think we said that, you

know, the environment is the foundation

and um security is also the foundation,

but the environment is really like um

it's like the foundation to both, right?

It's like it's also your governance. So,

um it's the architectural foundation,

but it's also the foundation of the of

the governance. So, it's almost like two

in one. If you have the environment

figured out, you also have like made

strides on the security

um bit with a which I found interesting.

So, um

let's let's talk about security and pick

that up um really quickly. So,

um you yesterday, if you tuned in, and

if you haven't, we spoke about um how to

actually avoid that agents are reasoning

around around the guardrails, basically

nice prompts that you put in to to ask

them not to execute certain things. And

um we had Stephen and Nuno from the open

source side talking about kernel level

enforcement. We had Leo Lorenzo and

Chris from Ona talking about VDO, the

data wall, and and how to actually um

protect at runtime at the kernel level

for for for that. And what I just found

nice at from the Monzo talk with Suhaib

Ellis

I I wrote I wrote this down cuz I found

it nice. Like he had a quote that said,

"There's nothing in regulation that

stops you from adopting an AI." And it's

exactly it's exactly the right mindset,

right? Like if you have those

foundations in place, if you can be sure

that you like limit your blast radius,

then there's like nothing stopping you.

And I think

especially we work a lot with the

regulated industries and oftentimes what

we hear is they they really believe they

can't use AI securely, and it's just not

true. Like of course you have to like

think a little bit different, and like

you have to do some of the protections

in place, but as you said, like if you

have the environment figured out, if you

have the governance of that in place,

then um you are actually very very well

set up.

>> Yeah.

>> And I also wouldn't

Yeah, go go ahead.

>> he said something too that I thought was

interesting in that last one, where he

said like people will always find a way,

you know?

Um

>> Yeah, yeah, yeah. Exactly.

>> Positive enforcement is so much better

than negative reinforcement sometimes.

>> E- exactly. I think shadow IT has never

been um higher than it is right now. So

even as a company, if you're like

I don't want AI cuz it's like it's bad,

it's too dangerous, whatever, your

engineers are one CLI command away from

from using those agents, and they will,

right? Um it's also what made cloud code

etc. like boom so much cuz it was like

uh such a bottoms-up bottoms-up motion.

So many of these companies are now in

the

um actually figuring out like how to

govern the the use of all those

different local agents. So

Yeah, really really interesting.

All right, let's move on. What's the

what's the next takeaway? I I that one.

>> Yeah, so the the third pillar when we

were kind of going into this, the there

were the three pillars, right?

Infrastructure, uh governance, and

leadership. So, leadership is where we

were trying to focus most of the talks

today, and I think a lot of really

interesting stuff

kind of came up. Um so, what really

stood out to you from some of these

talks around how people are uh how

people are rolling things out today, how

they're thinking about um enablement

beyond trust give everyone access to

background agents as a tool.

>> Um

yeah, I think there were a couple. So,

the

what we heard is that you you have to

you have to win everybody's trust. Like,

you have to start small. Like, none of

them came in with like, "By the way,

here's a big change. Tomorrow, your

whole workflow has is is is changing,

and like, please follow these steps."

Um I think the developers have been

following certain practices, obviously,

for for a long, long time. And um what

all these talks had in common is that

they're like really starting incremental

and small. And there were a couple of

talks that um um that talked about this.

I think Nico at Uber was was pretty

outspoken about this, where he said

like, I think he he was literally the

guy who said like, "Go where the pain

lives," right? And then it's like, I

think that's the that's the best proxy

you can you can do. Like, if you uh

identify the toil from developers and

can take that away, like, you earned the

trust. And then you go from there, and

then you go a little bit further. Um I

think Rajesh at Cloudflare is the same,

right? He talked about the PR reviews

that um created a lot of, you know, slop

and like slapped a bunch of comments on

the PRs. It didn't really help guiding

humans' attentions.

Um and then they changed uh to um to

their Codex approach. So, like, a

machine-readable, more deterministic

standard um that they would incorporate

into the agents' workflow when reviewing

the PRs. And like, suddenly, the um the

developers had trust into that standard

cuz they were used to it from the human

review, then they trusted their agent,

and then obviously that kind of resolved

the the the bottleneck. And

um the last one that comes to mind is is

Quake at Genentech where he literally

said, you know, they started with ops,

then they kind of graduated their agents

to also work on dev, um but then they

let the agent work on their on its

self-evolution, and kind of each of

these stages built the trust for the

next stage, and really gave them an

unlock. And now I think he he said

something like they have four a team of

four engineers, um and they have shipped

like 190 tickets or so in the last

sprint. So,

uh yeah. I think that's that's the the

gist of it is like start small, one

problem at a time, and then like, you

know, build your kind of factory up. Uh

as a whole thing.

>> really interesting to kind of take um

a You kind of have to take

simultaneously a prod like a platform

approach to solving this problem, but

also a product approach.

Right? Hearing the stories from like

Florence and Harvey about how they AB

tested the way that people interact with

the agent through things like Slack

summaries, and then how those Slack

summaries were presented to users versus

how they were presented to agents, and

the importance of like removing slop

from the potential outputs to improve

the trust,

um while also improving transparency.

There's so many things to consider in

the design aspects of a successful

platform here that you need to take a

product approach in building it. I

thought that was really really

interesting.

>> Great.

All right, what else do we got?

>> Yeah, I think at this point we've gone

through obviously 14 talks, and there

seems to be like one big theme that

feels like it's emerging.

And a lot of this is just best

practices, right? I mean, 20 years of

best practices in engineering and

platform engineering in agile

development, all of this just became

table stakes.

Right? If you want to start building the

future with background agents and

software factories, it's not nice to

have clear standards and the measurement

and service catalogs and IDPs, you have

to have them. It's a It is a requirement

moving forward in order for these agents

to behave

uh super well. I think someone uh I'm

blinking on who it was uh on day one. I

wish I had I wish I had the name in

front of me, but somebody mentioned um

like

I think it I think it was Patrick at

Tessell mentioned that every AI engineer

or every AI agent is effectively an

engineer that you're onboarding on day

one. And they're only as smart as the

things that you give them access to

instantly. Right? So,

it's more than just self-service. It's

more than just documentation and

onboarding. It's like you want to extend

the brain's capabilities effectively.

Um and that that's kind of the thing

that really was screaming screaming to

me throughout this entire summit.

>> Yeah. No, I I totally agree. And

um I get I think what all what what all

the others also had in common is that we

spoke about this a bit just now, but

like none of them started with this like

grandiose plan.

Um they earned the trust from there and

like going back to to to the

foundations, I think

um

it's really easy to get started with

that if you have the environment as a

foundation and if you um have the

governance around this cuz then

experimenting experimenting is actually

quite cheap, right? The blast radius is

like kind of limited. Uh you can run

things in parallel. Obviously, you can

run things in the background cuz they're

running in the cloud and don't have your

laptop

um to be open. And um you can experiment

with workflows one one workflow at a

time. Like can you can you know

continuously remove yourself as as the

bottleneck. I'm picking this up because

in the keynote we spoke about the

uncertainty increasing

cuz it can be daunting to like

and read like the end state of what the

the companies have been presenting here

like Stripe and Ramp etc. Like this is

the machine that we have and then you

look at this and you're like I'm

still stuck with cloud code in the CLI.

Like how do I get there? Oh my god like

they built so much for such a large

company I can never get there.

And I hope that these talks and like the

the journey that they all so kind of

also vulnerably shared is

made it much more approachable to to

everyone here and makes this again more

look like a journey rather than this

like grand grandiose blueprint that you

that you have to figure out.

So

the fastest way if you want to get

started if you're interested in that is

actually to abstract a lot of this away.

The thing you can abstract away is the

infrastructure piece and the architect

architecture around it. If you're really

interested in building all those

different layers that go into that that

some of those companies have spoken

about, you can check out

backgroundagents.com/landscape

where we collated 11 layers of the stack

with 95 tools that can be helpful to

building this.

Not many companies have the capacity to

to do it or

actually want to do it cuz it diverts

resources, right? So if you want to get

started right away with this

foundational piece in place, then

shameless plug this is exactly what Mona

has built for. It's got the obviously

the sandboxes context and security piece

and the adoption playbooks and

it really allows you to have background

agents execute end-to-end

on reproducible dev environments that is

that are connected to to to your whole

tool chain.

And

it also has these automations that we've

heard a couple of times today that for

example, Zach at the software factory

referred to. So, these are agent fleets

that are, um, running on your code base

or across thousands of repositories on

different triggers, like, uh, schedules

or web hooks.

And, um, all of that with the governance

and the guarantees that you need. So, if

you want, it runs in your VPC with audit

trails, scope credentials, and the

kernel level enforcement that we spoke

about in the in the talk with Chris Liu

and

Lorenzo.

So, if you want to get started, um, very

quickly with this,

um,

lock yourself out. We have a free tier.

Uh, go sign up and check it out. If you

if you require a little bit more

enterprise solution with running this in

your in your own VPC, please, um, don't

be a stranger and feel free to, um,

reach out to us. And again, if you want

to build the whole thing yourself, go to

the landscape page on the background

agent summit page.

Yeah.

Great. Will, anything to add on your end

here?

>> Um, no, I just wanted to kind of dive

into a bit of a thank you. Uh,

first things first, thank you to all of

our speakers. 14 incredible talks.

Um, every single person really put their

their heart and soul and all their all

their brain power into these

presentations for you guys. Um, we will

be following up with, uh,

we'll be following up with, uh, unlisted

YouTube playlists for all of you guys

have access to these recording uh, the

recordings of these talks if you ever

want to go back and reference them. As

Philip mentioned as well, we'll be

sending out a, uh, notebook LM, uh,

notebook LM podcast, um, and some other

useful tools if you want to go in digest

the content of these talks a little more

on your own terms and your own

timelines. Um, but seriously, we cannot

thank these people enough. Uh, it's the

research they've been doing that's made

any of this possible. It's the way

they've been implementing these these

pieces of technology in their companies

that made of possible and um, we're just

overjoyed with the quality of these

talks.

>> Yeah.

>> And on the flip side of that, I wanted

to also thank you send out a thank you

to everyone behind the scenes at Ona who

helped make this happen. So, Lou, if you

don't know who Lou is, go follow him on

Twitter at Lou JB. I'll send out his

handle.

He's kind of the He's the architect

behind the background agents landscape

and the micro site. He's also the person

that you saw running a lot of these

interviews.

He's extremely bright and understands

this space very very well. If you have

any thoughts, considerations, any

questions, just tag him on Twitter and

he will respond or X the everything app.

Likewise, Matt, huge thank you. He is

the

He is our head of engineering here at

Ona and he ran a bunch of the interviews

as well. You definitely saw his face in

the most recent the Monzo video as well.

And then a shout-out to the team that

made this webinar possibility. Ben,

Janine, and Sidhan for

making the website live and being behind

the scenes with pulling together all the

videos, editing, and graphics, and

everything. There has been so many

moving pieces to make this all happen.

So, just an enormous thank you.

>> Last one.

Thank you to you, Will, for for being on

with me here and like hosting through

all of this. Really appreciate it. And

lastly, thank you to the audience for

sticking around. It's now approaching 3

and 1/2 hours. We still have over 400

people watching live. Yesterday we were

even higher than that. So, really really

appreciate the engagement that we had in

the chat. Again, we will also follow up

with some of the answers from the

from the speakers that we weren't

answering live here. And

I think one one quick note, we will also

send around a survey. You will find this

also on the live embed page that you're

probably on right now. We would really

appreciate it if you give some feedback

on what um the quality of the talks

were, what we can do better, and if

you're also interested in um doing

something like that in person, live.

Not virtually.

>> Yeah.

Absolutely. Uh stay up-to-date with us.

Follow us on the Background Agents

microsite. Make sure to sign up for the

newsletter. Um cannot shout that out

enough. We put a lot of effort and a lot

of research into these. Um we will be

updating the website, we'll be updating

the landscape, and we'll be we'll be

sending out continuous, you know, actual

thoughtful emails. No uh no marketing

spam. Do not worry about that. So,

um I guess the final call to action is

we've we've learned about the

primitives, we've learned about how to

build this all out, learned about the

architecture, and the rollout and the

enablement best practices. So, with all

of that in place,

take this take this foundation and go

build it.

>> Let's talk about that.

Yes. And then, talk about us about this

in the next next virtual summit that

we're doing, or a live summit, who

knows.

>> Yeah.

>> Cool. All right.

It's a wrap. Thanks, everyone.

Take care.