- 28 minutes ago
AI agents are moving beyond demos and into the everyday workflows where companies design, build, ship, and operate. In this conversation, Thibault “Tibo” Sottiaux, Head of Core Product and Platform at OpenAI, and Peter Steinberger, creator of OpenClaw and part of the Codex team, will discuss how ChatGPT and agents like Codex are evolving from software development tools into a new layer of intelligence for work on the computer. From the builder community to enterprise adoption, they’ll explore what happens when software can reason across context, take action, and collaborate with people on complex tasks, and what it will take to make these systems useful, trustworthy, and broadly adopted across businesses.
Category
🤖
TechTranscript
00:26Hi, hi Vivacek.
00:28I'm so glad to be here.
00:29Once again, 10th year's anniversary.
00:32So, very glad to have you both on stage to talk about Agentic AI, the new busy world.
00:40So, I'm glad to have with me Thibaut Sotio.
00:43You are the head of a core product at OpenAI.
00:46And we have Peter Steinberger, the founder of OpenClo.
00:51Peter, first question.
00:52You joined OpenAI, it was in February.
00:55Yeah.
00:55So, it feels like an eternity in the tech industry.
01:00So, how did the onboarding go?
01:04Oh, you know, I think there's the onboarding for the usual path.
01:11But I was kind of just thrown into it.
01:14And it was just too much to do.
01:16And the first three days, I discovered our Slack.
01:20And it's like its own universe.
01:24And then I discovered that I can hook it up to Codex.
01:26And I can ask Codex all the questions.
01:28And I was like, oh my God, this is the best thing.
01:31But no, it just kept going.
01:36We also had early conversations of like, hey, you know, can I share this one on Twitter?
01:41Or like, is this one still secret?
01:43Oh, that's still a thing.
01:46So, Thibaut, last year we were mainly talking about, you know, generative AI here.
01:52Of course, at VideoTech, but all around the world.
01:53But now we are already talking about agentic AI, you know.
01:59So, what really changed, you know, in terms of tech?
02:03A bunch of things.
02:05The models improved.
02:06So, the models were capable, you know, suddenly of dealing with like more and more context,
02:10using tools efficiently, and then running for longer and longer.
02:15So, we saw this with GPT-5, you know, as well.
02:18And then like GPT-5.2 did another step change.
02:21And then combined with a different approach to how we were building the harness around it
02:27to just really let the model be in control.
02:29And that's how you get an agent.
02:30A very good model, a good harness.
02:32You combine it, and suddenly it can take actions in its world.
02:36Peter?
02:37You know, when there's this one picture from me where I sit there with 20 terminals.
02:43And in the beginning, that was the only thing how it would work.
02:47You would like talk to an agent, and it would do the work.
02:50And now, the model's got so good that basically you can talk to the model,
02:57and it will delegate the work to agents.
03:01You know, it's almost like we went a level up to talk to AI,
03:05and now we went up another level where AI talks to AI.
03:08So, we saw a lot of AI agents dedicated to software engineers.
03:16And now, we're seeing more and more agents dedicated to everyday work for business people,
03:23financial people, marketing people.
03:24So, what companies are starting to really use agents' AI?
03:31Are they still in pilot phases, or is it more?
03:36To take a step back, why do agents work for everyday work and not just for coding?
03:41It's like, if you think about it, what makes an agent really good at coding?
03:45It's understanding the broad context, the requirements,
03:49and then just doing, like, very goal-oriented work for a very long time with precision.
03:55And that is not specific to coding.
03:57It's, like, obviously very useful for coding.
03:59And in coding, you can always verify, usually,
04:01that the solution is correct a little bit more easily.
04:05But broadly, when you train the models for these kinds of capabilities,
04:08you end up with something, like, much more general.
04:11And what we were seeing on Codex is we wanted to bring additional tools,
04:15additional context, like, you know, from, let's say, Google Docs, Notion, Slack,
04:18and we were sort of, like, integrating all of the tools that we were using day-to-day.
04:23And at the end, we were realizing, oh, hey, like, people at OpenAI
04:27are using this for everything, you know, like, raising new fundraisers.
04:31Like, the last fundraise by Sarah Friars was, like, done with the help of Codex.
04:36And it was just, like, moving funds around and keeping track of things.
04:39And so we were seeing this mass adoption within the company,
04:42and we're, like, oh, you know, like, let's just adapt things a little bit
04:45and make this more generally available.
04:47And, like, we were, I would say, in the early stages of this,
04:50like, maybe a month, two months ago,
04:52and now it's starting to go, like, mainstream, I would say.
04:54I'll give you one example.
04:56I have a friend in the finance industry,
05:00and he always knew that, like, for the last year,
05:03I use AI and I do programming.
05:06And then one day he was just, like, watching me how I work,
05:11and it was not about programming at all.
05:13You know, it was, like, organizing my emails
05:15and, like, do random stuff with my computer.
05:18And then a few days later, he texted me that, like, he was mad at me
05:22that I didn't push him to try this out earlier
05:26because he thought it's just for coding,
05:28but now, like, he's using it all the time
05:31for all of his finance stuff
05:32because it's actually good for everything now.
05:36And designing an agent, you know, for code, of course, is one thing,
05:40but designing an agent for day-to-day work
05:43must be more difficult
05:45because, you know, human is unpredictable.
05:47So what are the main challenges behind that?
05:53You know, when you build,
05:58when you train the model to be good at coding,
06:01that is very easy, verifiable,
06:03because ultimately code you can compile.
06:06Either it works or it doesn't work.
06:08But for things like make a great website
06:11or write me a great speech,
06:13that's much harder to quantify.
06:16You know, what's good?
06:17Like, that's taste in the end.
06:21And that just was a lot harder
06:26to teach the model taste.
06:30And I would say that's still one thing
06:33where humans are just better
06:35and, dare I say, probably will always be better
06:40because ultimately the models are just very, very powerful tool.
06:44You know, it's like almost like a source hammer.
06:46You know, but the hammer alone is useless.
06:49You know, you still need someone that deals it.
06:53So...
06:55We build agents to help, you know,
06:58all of humanity at OpenAI.
07:00And one of the challenges
07:01with bringing coding agents to the world
07:03is, you know, solving the safety
07:06and alignment problem as well,
07:07where if you're giving this kind of, you know,
07:10PowerVille agent to someone who is not technical
07:12and suddenly it starts writing code
07:13and executing it on your computer,
07:15you're like, is that, should it do that?
07:17Is it doing something that's dangerous?
07:20Is it deleting my files?
07:21And so working on the security of it
07:24and being able to, you know,
07:26have very simple and verifiable controls
07:28so that, you know, you're like always in control.
07:31Like, you're not taking more risks
07:32than you're willing to take,
07:33even if you don't actually understand necessarily
07:36how the agent is operating under the hood.
07:38And we spend a lot of energy there.
07:40So today, many companies are deploying, you know,
07:43individual co-pilots, these co-agents.
07:46So at what point do we transition to an organization
07:51where multiple agents collaborate together?
07:56I would say, and Peter is going to have
07:57a different answer here
07:58because he's already, like,
08:00at the very late stages of agent adoption.
08:02But I would say it's, you adopt it piece by piece.
08:06So first you have, you know,
08:08maybe your own agent that helps you.
08:11You know, you just use it by yourself.
08:12You communicate with this agent.
08:14It is, you know, helping you try your emails,
08:17setting up filters.
08:18It is helping you, you know,
08:19prepare for an interview, these kinds of things.
08:22And then later you're like, okay,
08:24like there's a workflow that I'm doing every day
08:27or I'm doing every week.
08:28Let's set up an automation.
08:30And now suddenly it's running, you know,
08:32in a recurring fashion.
08:32It's running every day.
08:33It's running every week.
08:34And now you have an agent
08:35that's just, like, doing work for you
08:37in the background
08:37without you having to think about it.
08:39And then more and more of your team,
08:41like, start to adopt, like, the same patterns.
08:43And then you learn from that.
08:44You continue to automate.
08:46And then eventually they will just, like,
08:47talk to each other
08:48and act like as little teams
08:49and get steered by individuals at a company.
08:53And this is just really, you know,
08:54like going from very simple early adoption
08:58of, like, you know, just talk to your agent,
09:00get it to do things for you.
09:02Ultimately, you know,
09:02you have virtual teams doing work for the company.
09:05You know, sometimes it's also just useful
09:07to separate them a little bit by context.
09:11Like, I have my personal claw
09:13that, like, knows everything about me,
09:15but I'd also have one for my house.
09:18You know, I still have a place in Vienna.
09:20And everyone's just responsible
09:21for checking the cameras,
09:25checking all the KNX networking data,
09:27sometimes a prank to my cleaning lady.
09:32And then, of course,
09:33like, I have, like, one at work
09:35for specific projects.
09:38And at least with the tech we have now,
09:41if you specialize them,
09:42you usually get by far better results.
09:46Like, we're working towards this, like,
09:49one agent can do everything,
09:52but it's still very useful
09:54to, like, specialize.
09:56If you give the model more context
09:59and a narrower path,
10:02it will be more predictable
10:03and you get better results.
10:05So you think we'll have
10:06just specialized agent for a lot of things,
10:09not just a big one?
10:11But ultimately, also,
10:13it's still so early, you know?
10:15There's, like, a lot that we need to figure out.
10:16But I'm sure in the future
10:18there's going to be more than one agent.
10:20But where we land
10:22and how we structure things,
10:23I think nobody has figured out yet.
10:26So, of course,
10:28I'm sure a lot of people
10:29have agents in the room.
10:31But what about the adoption
10:33in the, you know,
10:34the more corporate, you know,
10:36corporate businesses,
10:37the traditional companies,
10:39if I may say that?
10:42Adoption has been exploding.
10:43And especially in France, actually,
10:45it was, like, looking at the stats,
10:46like, codex adoption
10:47has increased ninefold this year.
10:50We're seeing the biggest growth,
10:52actually, not on developers,
10:54but on, you know,
10:55just for day-to-day general-purpose work,
10:58which was quite surprising.
11:00I thought we were still earlier than that,
11:02but it seems like it is really going mainstream.
11:05And there's, like, since GPT 5.5,
11:08like, we've had so much demand,
11:09you know, especially from enterprises.
11:12Like, before that,
11:13chat GPT was, like,
11:14already, like, mass consumer.
11:15Like, we're approaching a billion,
11:17you know, like, weekly active users,
11:19which we will celebrate broadly at some point.
11:22But now, it's, like,
11:24it's going from conversation
11:25to everyone having an agent.
11:26And our personal mission
11:28is to just bring, like,
11:30a personal AGI to everyone
11:31in their pocket
11:33that knows and understands them
11:34and helps you get stuff done,
11:36just whether it is in your personal life
11:38or at work.
11:41I think you said everything.
11:44And is the adoption
11:45is better in the U.S.
11:47than in Europe?
11:49What is it?
11:49The adoption.
11:51Is it stronger here?
11:52No, the adoption
11:54has been growing faster in Europe.
11:56So we're not that bad in Europe.
11:58We're doing great.
11:59Oh, okay, good.
12:01I say we because I'm from Europe,
12:02he's from Europe.
12:03He's Belgium.
12:04Everyone's from Europe here.
12:06We are a lot of about,
12:08so we talked about
12:09a lot of the capabilities
12:11of AI agents,
12:12but I want to talk about,
12:13of course,
12:13the limitations.
12:15What is the main obstacle today,
12:18you would say,
12:18to this adoption?
12:19Is it, I don't know,
12:21technology,
12:22governance,
12:23cybersecurity,
12:24or just humans themselves?
12:26Honestly,
12:27the gap between
12:29what we do with those models
12:31and what they are able to do
12:34has never been higher.
12:35So I always say,
12:37like, the main issue
12:39we have right now
12:40is imagination.
12:43Like,
12:45there's still so much
12:46we can build,
12:47there's so much
12:48we can make,
12:49better with more intelligence,
12:50there's so much
12:51where we can build
12:53clever new tools
12:54around AI
12:54to help us
12:55in our daily lives.
12:57You know,
12:57like,
12:57even a tool like OpenClaw,
12:59somebody could have built it
13:01three, four, five,
13:03maybe six months
13:04before me,
13:04but it somehow
13:06didn't happen.
13:07Who knows
13:07what else
13:08people can come up with.
13:09And I think
13:10that's also the power
13:11of AI,
13:12that you cannot
13:15just prompt things
13:16into existence,
13:17even if you
13:18don't exactly know
13:19how everything works
13:20in detail,
13:22the agent is
13:23infinitely patient
13:24and you can just
13:25ask whatever you need.
13:28To add to this,
13:29I think to me
13:30there are like three things.
13:31Connecting to the right
13:32context and tools,
13:33which takes time.
13:35You know,
13:35this needs to be integrated
13:36within the company.
13:37It is not something
13:38that can be automated yet.
13:40Security,
13:40making sure that you're
13:41not taking unnecessary risk
13:42and can always,
13:43like, you know,
13:44have the right observability
13:45and controls.
13:46And we're building
13:47a lot of things there.
13:47For example,
13:48we're innovating on
13:49agents monitoring
13:50other agents,
13:52such that we always
13:53verify that the intent
13:54of the user
13:55is respected
13:56and we block
13:57certain actions
13:57if, you know,
13:58suddenly it's like
13:58starting to do
14:00something higher risk.
14:01And the last one,
14:02which is like Peter
14:02talked a lot about,
14:03is like education, right?
14:04This, like,
14:05breaking free
14:06from your previous
14:07patterns and habits
14:08and understanding
14:09that now suddenly
14:10you can do more.
14:11It is,
14:12if you set the right
14:13objective,
14:14if you explain
14:14what you want
14:15the agent to do,
14:16like, more often than not
14:17it will be able to do it.
14:18And, you know,
14:19every time it surprises me
14:20and, like, you know,
14:20with each model generation
14:22we're like, oh my God,
14:22it's like it's doing
14:23way more than I expected.
14:25But you know what I also
14:25find so useful is
14:28just sitting with
14:29other people
14:29and watch them
14:30how to use their agent.
14:32You know, there's like
14:32sometimes I watch people
14:33and I'm like,
14:34yeah, I kind of understand
14:35why you're not getting
14:36good results.
14:37And then other times
14:38I watch people at the company
14:39and I'm like,
14:40it can do that.
14:44I want to talk
14:45a little bit
14:46about the future.
14:47But the future,
14:48it's maybe three months,
14:49you know,
14:50in this industry
14:50because things
14:51are going so fast.
14:54If agents
14:55becomes, you know,
14:56the primary interface,
14:58what will the,
14:59just a SaaS,
15:01what will it become
15:02in the next few months?
15:03It will disappear.
15:05We've talked a lot
15:06about the SaaS apocalypse.
15:08It didn't happen
15:09yet,
15:10but I want to know
15:12your opinion on that.
15:16I think it's going
15:16to get bigger.
15:18You know,
15:19yes,
15:20you can type
15:22in a small prompt
15:23and you get
15:24something out.
15:26And that's going
15:27to be a big use case,
15:28like this quick,
15:29disposable,
15:30useful software.
15:30But if you want
15:32a reliable service,
15:34there's a lot more
15:35that goes into it.
15:37Like,
15:37software doesn't end
15:38at creating it.
15:40Software needs
15:40to be maintained
15:41and that's a lot
15:42of work.
15:43Like,
15:44many times,
15:46people,
15:47it'll still be
15:48more economically
15:49if people pay for it
15:50than if they build it
15:53and also have to carry
15:54all the maintenance.
15:55I think we're going
15:56to continue to see
15:57deep infrastructure
15:59and a lot of investment
16:00there and that's not
16:01going to go away.
16:03You know,
16:03like the databases,
16:06like back office,
16:07like,
16:07you know,
16:07it's going to,
16:08I think,
16:08continue to be standardized
16:10and heavily invested in.
16:12Maybe not in three months,
16:13but at some point,
16:15I do think everyone
16:17is just going to have
16:18their personal agent
16:19that understands
16:20your goals
16:21and your preferences
16:22better than any other software
16:24and that,
16:25the agent
16:25and the model
16:26behind the scenes
16:27will be able to generate
16:29software on the fly
16:30and you'll have
16:31your very own
16:31personal software,
16:33personal dashboards,
16:34and it will all
16:35be tailored for you.
16:36And at that point,
16:37do you need to interface
16:38with anything else
16:39other than your agent?
16:41Probably not.
16:42One thing that I like
16:43to think about as well
16:44is humans don't have a screen.
16:45Like,
16:46you know,
16:46you can't,
16:46you know,
16:47see what's happening
16:48in my brain.
16:48Like,
16:49I don't display,
16:50you know,
16:50I'm just like gesturing.
16:51Like,
16:51you know,
16:51I'm in physical space
16:52with all of you.
16:54We speak to each other.
16:56And,
16:57you know,
16:57maybe agents in the future
16:59will act in a very similar way
17:00where,
17:01you know,
17:01they will just be like
17:02in your natural environment
17:04through natural conversation
17:06instead of,
17:07you know,
17:07being restricted
17:08on just like your phone.
17:10Everyone will be a CEO.
17:14Everyone's a manager,
17:15in a way.
17:16Everyone will get
17:17tailored help
17:18based on like
17:18what they're trying to do.
17:19Yeah.
17:20Okay.
17:21A question.
17:22Which company function
17:23will be the first
17:24to be primarily
17:25carried out
17:26by agents
17:27according to you?
17:29Yeah,
17:30I have a very
17:31specific question.
17:33The first?
17:34What was the first thing
17:36that was ever done
17:37by an agent,
17:38do you think?
17:42I mean,
17:43what helps me a lot
17:44is using agents
17:47as an assistant,
17:49as a personal assistant.
17:50Like,
17:51give me a briefing
17:52of my day
17:53when I have a meeting
17:54with someone,
17:56write my introduction,
17:56who is this person,
17:57what do they do?
17:58Agents are so good
17:59at like using
18:00all those data sources
18:01and connectors
18:02and then compiling it
18:03into something
18:04that I can quickly read
18:06in a few minutes.
18:07But I wouldn't say
18:10the job of that person
18:11necessarily goes away.
18:13It's more like
18:14they will probably
18:15compile it
18:16than for me
18:17and they have more time
18:21finding better sources
18:22or like helping me
18:23with other things
18:23on my job.
18:26What we were seeing
18:27for coding
18:27is like people
18:28like to use it
18:29to do things
18:29that they don't like to do.
18:31So, for example,
18:32I program every day.
18:33I don't like writing tests.
18:35You know,
18:35maybe my agent
18:35can write tests for me.
18:37And I think
18:37this is going to be
18:38the same thing,
18:39you know,
18:39for like everyday work.
18:40It's like
18:40what is the boring things
18:42that ought to be done
18:43but that you're not doing?
18:45Like perhaps
18:45they will start to be done
18:46by your agent initially
18:48and then you'll also realize
18:50like you can actually
18:51do even more
18:51and you can do
18:52like really cool things
18:53with it.
18:55So, we're at Vivitech.
18:56So, let's come back here
18:58in 20, 30,
19:01so four years.
19:02So, what will seem obvious
19:04to everyone by then
19:05but still seems so crazy today?
19:09You know,
19:09when I wrote this blog post
19:13about how I write code
19:18but don't read it,
19:19you know,
19:19how I just prompt it
19:21and ship code
19:22and it was
19:24at the end of December
19:25and people thought
19:26I'm crazy,
19:27you know,
19:27it's like,
19:28oh,
19:28it's just all slop
19:30and slowly
19:31the world
19:32is reaching
19:33an era
19:33where this is
19:35becoming the norm
19:36and if you think
19:37in a few years
19:39we're going to have
19:40a billion programmers
19:42that don't know
19:43that they program
19:44because all they do
19:45is they ask their agent
19:46for a solution
19:48to something
19:49and then they'll get
19:50their solution
19:50and maybe they get
19:51their personalized
19:52little game
19:53or a dashboard
19:54or a customized UI
19:56for faster learning
19:57or whatever
19:57they are wishing for
19:59without realizing
20:00that under the hood
20:01it's writing code
20:02because it's just,
20:04we just upped
20:05the level of abstraction.
20:08I think all of that
20:10and by 2030
20:12I think we'll become
20:13accustomed to
20:14like breakthroughs
20:15after breakthroughs
20:17in science
20:19in, you know,
20:20across like biology,
20:22chemistry,
20:23medicine.
20:24I don't think
20:25we'll go like a day
20:26without, you know,
20:27having like solved
20:27like another thing
20:28that we didn't solve
20:29for like 100 years
20:31before that.
20:32I think all combined
20:33agents will produce
20:34more value
20:35or, you know,
20:37maybe significantly
20:38more value
20:39than the entirety
20:39of humanity
20:40in 2020.
20:42And then also
20:43the way that we will
20:45interface with these agents
20:46will be extremely natural.
20:48It will not be through
20:49like a little problem box
20:51and having to think
20:52like very hard about it.
20:53They will just really
20:54truly understand
20:55what you're looking for.
20:58And I want to talk
20:59a little bit about Europe
21:01because you say
21:02you're from Belgium,
21:03you're from Austria,
21:04I'm from France.
21:06Is it a good opportunity
21:07for Europe
21:08what's happening right now
21:09with the AI agents
21:10and especially
21:11for European founders,
21:13European builders?
21:17I mean,
21:18if you look at Europe,
21:20like I love
21:21that we have Mistral.
21:22I hope in the future
21:23there's going to be
21:24more companies
21:25that are on the frontier.
21:28Ultimately,
21:29if there's more
21:30players in the market,
21:32it will push
21:32every zone
21:33to innovate faster.
21:34And I'm a builder
21:35at heart,
21:36so I love that.
21:39I think there's
21:39incredible energy
21:40in Europe
21:41to build.
21:43and do things
21:44that are useful.
21:46I often find,
21:48you know,
21:48whenever I look at
21:49startups and ideas
21:50in Europe,
21:51I'm like,
21:51oh,
21:51that's like deeply human.
21:54It's rooted
21:54in something real.
21:55And I think harnessing
21:56that and,
21:57you know,
21:58continuing to innovate,
21:59you know,
21:59just being rooted
22:00in solving real problems
22:01for humans
22:03is awesome.
22:04Like,
22:04we invest a ton
22:06in the open ecosystem.
22:07A lot of our code,
22:08you know,
22:09both for OpenClaw
22:09and Codex
22:10is open source.
22:12It even supports
22:13alternative models.
22:15Yeah.
22:15You don't have to use
22:16OpenAI models
22:17if you want to use Codex.
22:18And I think
22:19we're going to see,
22:19like,
22:19a lot of different ideas
22:22and innovations,
22:23like,
22:23flourish from,
22:24like,
22:24all across the world.
22:25And maybe someday
22:26you will come back
22:27to Europe.
22:28Maybe.
22:29Yeah.
22:30I'm here all the time,
22:31so...
22:32We're here now.
22:34Well,
22:34thank you very much,
22:35both of you,
22:36for coming to Vitex this year.
Comments