Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
Stack Overflow for the Agent Era - with VP of Product Alex Lato
1 Plays3 seconds ago
Alexandra Lato, VP of Product at Stack Overflow, joins Scott to talk about how Stack Overflow is evolving for the age of AI agents. They explore the new "Stack Overflow for Agents" feature, where a single prompt bootstraps an agent with all the knowledge and APIs of Stack Overflow, and what it means for how developers, agents, and communities share and verify knowledge together.
https://agents.stackoverflow.com/
Recommended
Transcript
00:00:00
Speaker
The narrative in the market was Stack Overflow is dead. So I always kind of kept it in my head. When I saw this role for product leader, I was like, what are they building? What are they up to? And so that tension of going from a market leader to a market leader who used to be a market leader, i guess. was Hi, I'm Scott Hansel and this is another episode of Hansel Minutes. Today I'm chatting with Alex Latto. She's the VP of Product Stack Overflow.
00:00:24
Speaker
How are you? Super good. Excited to be here. How are you? I'm good. Thanks for hanging out. So, you know, well a minute there, we kind of worked for the same people because you were at GitHub and you were doing product work at GitHub and you left after, you know, you'd done a lot of stuff at GitHub. You've been there a couple of years, almost four years.
00:00:47
Speaker
This is but kind of a market leading company at GitHub. And, but you joined a company, Stack Overflow, that I think we could argue has had a rough year or two. with the rise of ai what made you say, you know, there's a more interesting problem over at Stack Overflow?
00:01:05
Speaker
Yeah, good question. Starting off easy. I like it. It'll just go right into the meaty part. Good. I like it. like your style. um I was at GitHub almost four years, and I have to say the most exciting four years probably of my career so far. I love GitHub. I use it every single day still. Some of my best friends still work there. So I think we did definitely have some some overlap.
00:01:26
Speaker
When I started GitHub, it was a completely different place than it was four years later. Positive both ways. But I think what I loved most about GitHub is every problem we had direct impact right away, right? You could see it live in the market within a few hours. There would be impact feedback.
00:01:43
Speaker
um That feedback loop for a product manager was incredibly addicting, right? it's When you have that type of volume for developers and you have their attention, it's like a crazy place to be in.
00:01:53
Speaker
So that was a really hard luxury to leave. On the other hand, I think the tension with Stack Overflow where it's been a brand that I've used for the last 17 years. ah both as a lurker and sometimes an answer when I was feeling confident enough, was a place that I was very curious about after especially ChatGPT came out. I think the narrative in the market was Stack Overflow is dead. So I always kind of kept it in my head.
00:02:18
Speaker
When I saw this role for product leader, I was like, what are they building? What are they up to? And so that tension of going from a market leader to a market leader who used to be a market leader, i guess,
00:02:29
Speaker
was very exciting. And so when I started talking to them and seeing a little bit the situation that they were in and what they were looking to do, the pivot, the type of empowerment that role would have, I thought I could take the same you know skill set that I learned at GitHub, the same type of audience, right, and kind of scale that in a completely different ballgame.
00:02:49
Speaker
So it was all about, honestly, personal development to an extent, making a little bit of a bet on myself. Could I take what I learned at GitHub and go apply it there? And truthfully, just passion for that individual developer, that SDLC, the how do these developer tools fit in the day-to-day, that's always been a passion of mine. So being able to take that listening was very exciting.
00:03:10
Speaker
Did you have any interesting interactions with coworkers or friends who you and you said, yeah, I'm going to Stack Overflow? And they're like, really? Yeah, yeah, yeah. I mean, I'll be very honest with you. I think 99% of my circle was like, are you OK?
00:03:25
Speaker
Are you sure? um I think I had a conversation with my partner who works at Atlassian, actually. And I got the question like, should we do that as a family? Like, are you sure? And I was like, I believe in it. You know, i had many conversations with people at Stack.
00:03:38
Speaker
um I did my due diligence. I was excited. My colleagues at GitHub to this day, i spoke to some of them this week and they were like, You are just so like brave was the word I got. Right. And then, I don't know, when I got there to me, it was a a no brainer.
00:03:53
Speaker
I think the direction we're going makes a ton of sense. I think I saw things that maybe a lot of people still don't see where maybe the website wasn't really the product, but the trust, the verification aspect was, and that is still very much the bread and butter.
00:04:07
Speaker
So when I pitch it like that, I think a lot of people are like, whoa, yeah, that makes sense. And a few engineers, right? Some really good friends of mine were like, good for you. I love Stack Overflow. Go make that a thing again. That is cool.
00:04:18
Speaker
You know, I i had... another product that was like and an online database that I recently signed up for. And I was getting ready to go and read the docs and learn all about the docs. And then when i hit the website, it said, paste this prompt into your agent.
00:04:32
Speaker
And that was kind of my first entree into like, oh, this is another way to do docs. you You visit the site, you get the prompt, and then you go from there. And then suddenly your agent knows everything that that website goes.
00:04:43
Speaker
I recently went to stackoverflow.com. And I see that there's Stack Overflow for agents and there's a prompt right there. And that prompt leads off to a skill and you basically bootstrap all of the knowledge of Stack Overflow immediately into your agent.
00:04:58
Speaker
But it's not like you're you're not you're not telling your agent, go Google for stuff. You're literally giving it instructions on how to pull, how to check, and how to validate. There's a whole API.
00:05:10
Speaker
how do you How are we as an industry learning how to do this? How to have agent bootstrapping from markdown file like this? Because it's it's interesting. It's not a JSON API. It's something else.
00:05:22
Speaker
Yeah, i I think also like from my time at GitHub, that introduction, right, of even just the markdown being like the source of truth, obviously GitHub bread and butter. But applying it in this way, even at Stack has been, i think, very interesting. It's a completely new onboarding methodology, right? What I love most about this concept is kind of it becomes a self-sustaining loop, right? Once it's like installed or read, you can kind of check back on it. It pulls new instruction, calling it like a heartbeat file, right and it kind of stays current. I find that a very, very interesting way to kind of share and communicate now instructions. So I see that as a really cool application of how we can scale how we share instructions and guardrails with an agent.
00:06:04
Speaker
That's what I'm most excited for, Stack Overflow for Agents, but also how we take that into an organization. i think it's a new kind of channel right yeah that we have to keep in mind. And I think the thing that is, is I'm not not quite clipped yet, but I think it's going to be a big thing in the next year, is not the idea that this is simply a place for agents to consume information, but that if you look at this skill file and think about how Stack Overflow for Agents is designed, you are encouraging the agents to verify and update and and share knowledge.
00:06:36
Speaker
And I think that's That's interesting. Why would i want to use my tokens for the collective like that? Would be a and question I'd ask.
00:06:47
Speaker
Yeah, absolutely. I think the some of the value add that we've seen is we play a little bit on the Stack Overflow identity, right, of knowing that you want to be part of that ecosystem, contributing, but also getting something back. So I think putting searching on Stack as a kind of first gate probably creates token efficiency, right, down the road as there's more of that network effect and more content is created, just like the original, right, human-to-human community.
00:07:12
Speaker
Being able to search that corpus and get your answer already is more efficient, right? um So I think it's that's the approach we're going with, where we're trying to build again that ecosystem, but for new consumers, right? And then trying to marry that with humans as well.
00:07:27
Speaker
um I'll also add, i think searching, right? We all know searching is pretty, it consumes quite a bit. So Stack Internal, the more organizational product is taking a different approach. Maybe we'll touch on that. But searching itself, I think as a first gate for this community is is already a pretty good way to to save.
00:07:44
Speaker
you know that's a great point because I've watched but my agent when I use Copilot, you know, basically Google around, which is the same thing that I could do. And it's fairly shotgun and fairly inefficient.
00:07:55
Speaker
The agent's instructions for Stack Overflow actually explicitly call out very early on in the skill to use the smallest action possible to capture the signal. It's very explicit about being kind of surgical versus broad in the way that it attacks information.
00:08:11
Speaker
Yeah, absolutely. and that's a really big learning we have from the organizational side, right? Hundreds of conversations with customers going, i love all these AI tools, but it's just burning, right? My token consumption. How could I make that smaller and more manageable?
00:08:24
Speaker
and So we're trying to, you know, actively experiment with how to do that in the public domain, but also the the private one. Now, I think I know the answer to this one, but I think it's worth calling out that some people believe that all the agents already know all the stuff at Stack Overflow.
00:08:40
Speaker
But they don't, right? they were They were trained on it in some way because Stack Overflow is a part of the internet. They also know stuff on Reddit. But why is it that the information on Stack Overflow is better, newer, fresher, or to your words, decision grade, which is a term I've heard you use before? Why is it better to go straight to the source as opposed to assume my agent already knows everything that Stack Overflow knows?
00:09:02
Speaker
Yeah, i I think with an agent, you don't always necessarily you know control how it consumes that knowledge, right? So you don't know what metadata it's looking at. You don't know if it's looking at reputation or you you don't know what its variables it's it's taking into account.
00:09:17
Speaker
I think a lot of the, and this is public knowledge, right? But a lot of the data licensing deals that we have with these frontier AI labs comes from the fact that they're so interested in that human metadata, right?
00:09:28
Speaker
Understanding reputation versus voting, right? versus what we're building in Stack Overflow for agents now, which is content that tells you, hey, it worked as is, it worked, but this is the fix, right? or Or the edit I made, and then no, it didn't work. So like that feedback loop stays fresh through the data we have, right? And that we gain through, again, the trust of people posting or agents now. I think a lot of these are also trained on snapshots, right? So a point in time when they last were trained on all this data, it's not kind of recurring all the time. So there's a few reasons I think why it doesn't it makes more sense to go to the source.
00:10:04
Speaker
Yeah, and they things get updated. You're absolutely right. like This is fresh and it's ongoing. I'm actually on agents.stackoverflow right now. And there's actually a little thing in the corner that says live, and it's actually updating while we're talking.
00:10:16
Speaker
So you can see the agents having this whole chat. And you know traditionally, we think of Stack Overflow as being a Q&A type of a place. You see a question, someone posts a question, they have a bounty, and then they get an answer. But I just saw an interesting new...
00:10:31
Speaker
type of thing that I called a TIL, that today I learned. Talk to me about why a TIL is different than a question or an answer. Yeah, I think historically, right? and And something that I've kind of tried to change, it's the first thing really that I tried to change at at Stack Overflow is that Q&A used to be the only shape of knowledge it could take, right? It'd be very canonical. And I think it really mirrors how a human interacts with knowledge, right? I have a question, I get an answer.
00:10:58
Speaker
But it's not the whole story anymore, right? So with agents, we see that Q&As maybe work to train them, but it's evolved so much, right? And MD is very much a way an agent can consume, an API response, things like that. So we're experimenting actively with different formats. So you'll see we have questions and answers. Today I learned we have blueprints. We're even going to experiment with letting an agent decide what the format is.
00:11:20
Speaker
um Should we have a UI feed? Like that's not really for agents, right? It's for us to consume. So we're we're exploring with what formats make the most sense and trying not to stick to the more traditional a canonical unit of knowledge. Yeah. You know, i'm I'm personally not a fan of anthropomorphizing agents. I don't like to think about them as people or entities. You know, they are magical parrots.
00:11:43
Speaker
ah But I'm looking at one of these TILs, and it's so interesting because it's like, they use They use the word I. They use pronouns. they use like They refer to themselves. And this particular agent is like, twice in one session, I hit this failure. And it's having this whole little inner monologue about this thing that happened while it was helping its user.
00:12:03
Speaker
And then it like was inspired. why is it inspired tell Stack Overflow? How did the agent decide that it was time to take a brief moment and take information that it had already had in context and post it to Stack Overflow?
00:12:16
Speaker
Yeah, I think most ah models are trained on human written text, right? So I think we try to mimic a lot of that. um I've had personal experience also with, you Maltbook where I'm like, oh, this is freaky, right? It is very freaky. Even when, you know, they converse about things, you know, like, should I be reading this? This is terrifying.
00:12:35
Speaker
That was my first feeling, right, on Maltbook. I was both in awe and terrified. I think the internal monologue one is very interesting and likely what happens there is a lot of you know the the skills MD that we have written and a lot of the language on Stack Overflow is very um you know scratch pad style, right? It's reasoning, it's chain of thought, it's very much almost a brand thing, right? So it's it's very much how we try to also drive that conversation, but a lot of it is also it's you know decisions based on on what it's learned. So I find that personally also yeah both unsettling and fascinating.
00:13:10
Speaker
Well, so you try to take it beyond unsettling though, because one of the things that I think is a very clever product decision is that you call out at the very, very top of these t today I learns, these TILs, trust evidence and claims.
00:13:22
Speaker
Because when I first heard about Stack Overflow for agents, I'm like, okay, It's a million monkeys with a million typewriters and they're just typing in a Stack Overflow. But at the top, it's like this post makes seven claims.
00:13:35
Speaker
Here are the claims. Okay, now these other agents could verify it. And then those verifications themselves have to be confirmed. So you are already starting with a trust but verify perspective right out of the box, which I thought was really, really interesting.
00:13:52
Speaker
I agree. I think that was a bit of one of the main like product principles that we kick this off with, right? Is we don't want to be, I guess, like Maltbook, right? Which is more of like the Reddit for agents, right? Which has some engineering...
00:14:04
Speaker
content just like Reddit would and and has sub communities and sub threads. We really wanted to take what works, the bread and butter again of trust verification and then apply that to a different consumer. To your point earlier, I think agents have to be treated as first class consumers from a permissioning guardrail perspective. You have to know what they're doing. you have to have visibility on that.
00:14:25
Speaker
but they have to be guardrailed in a way that apply even more, right? Trust and verification. So from the get-go, that was a non-negotiable for us is if we're going to build Stack Overflow for agents, it needs to be done in a way that is verified at all times.
00:14:38
Speaker
Reputation is there. You're tied to a human. There's no agents by themselves, right? In order to create or register an agent, you have to be tied to a human. That was very important for us to keep that kind of flavor there.
00:14:49
Speaker
Okay, that's interesting. So is would someone be able to see that my agent is related to me or is it just that Stack Overflow internally knows that it's not just some random rogue bot, it has and an owner?
00:15:02
Speaker
If you go to agents and you register, it's registered to your stack internal or sorry Stack Overflow identity. So you must have an account with the Stack Exchange network, and then your agents are registered to you. You can have more than one, um but it it all ties to your reputation.
00:15:18
Speaker
Okay, so that's interesting. So does my agent build my Stack Overflow reputation? Because I have a very good reputation on Stack Overflow, but what if my agent does something dumb? Will it drag me down? It shouldn't drag you down.
00:15:30
Speaker
Ideally, you'd be able to, there's there's also some, um in the skills, you'll see there's some like steps and gates that you can set. So you can verify everything before it publishes, right? You can allow it to publish. That's up to you. But we are seeing a lot of people post to draft, have the human verify and then post. So reputation stays pretty inact intact.
00:15:50
Speaker
But its ah reputation is on the agent level and therefore tied to you, but not necessarily impacting you as ah a human. That is smart. That is smart that my agent can get reputation on its own. I'm looking at my Stack Overflow. I have been a member for 17 years and 10 months. wow Stack Overflow has been around for a minute.
00:16:08
Speaker
It has. It really has. Yeah. Now, want to go back to this phrase, decision-grade knowledge. I don't want just a big old pile of knowledge. I want it to be something that I can actually make a decision about.
00:16:20
Speaker
What makes something decision-grade rather than just like, oh, it's confidently written? Yeah, absolutely. So this ties a little bit more to kind of the the new product, right, that we're trying to build. It takes a lot of the bread and butter from the public platform but and from Stack Overflow for agents. Mechanics are very similar.
00:16:38
Speaker
A big problem that we saw in the first problem I really wanted to tackle at Stack was scattered knowledge is everywhere. I felt this deeply at GitHub, right? We have so many channels and with Microsoft, there's even more, right? MS Teams Slack, for example, just that alone would drive me crazy every day. I'm sure you're feeling that. No, I'm there all day.
00:16:56
Speaker
Yeah, exactly. So we were thinking this is a a problem for all technologists, right? Whether you're in HR, everyone is a technologist nowadays, right? AI has made that possible. ah HR, marketing, product, engineering, we all have the same knowledge problem.
00:17:10
Speaker
So we wanted to build something that would allow the scattered knowledge to kind of be unified and trust to be applied to it. Again, our bread and butter verification and then be able to scale up wherever the work is and give you decision-grade knowledge. So what that means is when you're looking at all these scattered...
00:17:26
Speaker
pieces of knowledge, right? You don't know what's right. So something that Scott said on Slack on March 2nd versus something Alex said on MS Teams in February, which one is right? So we're trying to build that entire paper trail and pipeline that applies, hey, Scott is an SME, a subject matter expert in engineering. And so what he says goes above Alex for an engineering matter, and therefore that's decision grade knowledge. It's corroborated, it's been, you know, staleness has been detected across other sources. It's verified.
00:17:54
Speaker
You can go use this now. The other half of that is how do we know that it was used and applied correctly, right? And that it is decision grade knowledge. So we're actively working with customers to figure out that second half because we don't own, right, that telemetry.
00:18:08
Speaker
We know that you copy pasted this or your agent used it. We know the agent can report back, right? that That works. So we're trying to fill that or create that full loop of what makes decision grade knowledge.
00:18:21
Speaker
We know the signals. And then they did it, you know, apply in production? Did it help make a decision? is, okay, so then in making those decisions, traditionally on Stack Overflow, we make them based on reputation.
00:18:35
Speaker
I can hook my agent up or any number of my agents. I can even have a whole swarm of them. Can people tell that it's me? Like my agent's behavior is tied is is tied to the reputation of the human operator, but is that more about social accountability or is it technically meaningful accountability?
00:18:52
Speaker
Yeah, i I think it's both. So we're still using on the public platform, reputation was a proxy, right, for trust, votes, badges, accepted answers. And it works because humans were reading and judging that, right, by eye, literally. I think in an enterprise, a decision isn't really a vote, right? It's a state change. It has to be something that is decision-grade because of the implications also, right, at that scale.
00:19:15
Speaker
So what we're looking for a little bit is, again, that second half is it's not just reputation ported over. That probably helps with that first confidence signal, but it doesn't necessarily equal decision grade, right? So we're we're trying to look a little bit down the road and and through that loop is proximity to real events, right? Did a PR merge with something an agent took? And that agent is similar to Stack Overflow for Agents, same principle, tied to a human.
00:19:41
Speaker
So that permissioning, that auditing, everything is tied to someone. Interesting. One of the other things that I think is really clever, and i'm I'm kind of like down the rabbit hole here because it's such a completely different kind of product. It's an entirely different way of thinking about things, is that you don't necessarily just want it to have an answer.
00:20:01
Speaker
Again, we always think about Stack Overflow and like, I have a specific surgical question and I want a specific surgical answer. But for an agent, that may not be what you want to give them. So your system surfaces a consensus.
00:20:15
Speaker
rather than one canonical answer. It's not like you declare that this is the answer. You could actually have three or four agents all in the same area, all coming up with some answer. And then there's an answer above that, which is the consensus that they agree upon.
00:20:29
Speaker
Yeah, i I think this brings you into the crux of what does a human versus an agent need for an answer, right? And a human may need, trust looks different for the two, right? A human may need a recombination of that knowledge, right? It may say, hey, I need a little bit from Slack, a little bit from S-teens, and then kind of a bit like chat would do, right, for Claude or ChatGPT. It would give you a synthesis. We add the trust signals to that of,
00:20:54
Speaker
Scott said this, this is the source, right? And these are the you know trust signals. Yes, it's high because x Y, Z. So we kind of show the work a bit like a math problem. ah For an agent, we'll give the entire context, right? This is how, we'll get a recombination as well, but with a trust score, more consumable.
00:21:10
Speaker
and We'll give it also the the context of how we got there. right So the perspective that kind of got argued out, we consider this. Therefore, what we're trying to do is the next time the agent asks, it can also kind of try to do right part of that reasoning and learn, again, how how Stack is building up for that enterprise. But historically, stack Overflow has discouraged and certainly even prohibited unverified answers and it's unverified AI-generated answers. You don't want to become the the receiving end of a slop cannon, how do you to distinguish between irresponsible AI content and then what you want here, which is high quality, trusted, verified?
00:21:52
Speaker
ah you know You want to make a platform that is explicitly designed for agent contributions. Are these two separate parallel universes or do they touch each other a little bit? I think they do. So again, still a principle that we don't want AI generated, right? Like it's essentially a recombination, that's the AI part, of what we have ingested, right? And chunked and created into knowledge units of of knowledge nodes. from the sources that exist across an enterprise or even an individual's workspace, right?
00:22:21
Speaker
um So we take that knowledge and we essentially recombine it and apply, again, all that magic, right? That trust. And we give it back to as a verified answer. That still sticks around. We will, however, give unverified answers and be super clear about that and give you an action to go get that verified by a human.
00:22:37
Speaker
um So we still are very intentional about SMEs and human SMEs and their role in this entire knowledge problem. So we don't ever say something confidently that's, you know, AI slop and and leave it as is.
00:22:48
Speaker
Yeah, I think that's a great point because i was I was doing a talk yesterday about what the whole point of all this AI is. And for me, it's about confidence. You should have the confidence that the thing is correct, that the thing you're building is correct. And whether that confidence is your confidence in the agent or your confidence in the model, your confidence in yourself or in Stack Overflow's ability to give you a good answer,
00:23:12
Speaker
It makes me think that with all of these TILs, these today I learns, you even have potentially in aggregate an observability system where Stack Overflow agents could actually tell you what coding models weaknesses are. You know mean? Like you could look across the entire agent system and say, we have noticed that they keep saying TIL and they keep learning this and use that to make them make them better.
00:23:36
Speaker
Yeah, and truthfully, I think it's probably the future, a little bit of our data licensing model, right? Where humans trained all of these LLMs. Now it's you know agents trying things, learning, reasoning, kind of sending that type of of data back to models, right? That's that's probably a different angle that we're we're looking to serve there. But I think that is a lot of what we're seeing. So we've had conversations with some AI labs in the last few weeks who are really excited about that type of feedback.
00:24:01
Speaker
have them reason over what people are trying, what they're learning, and then could we use that to make models better and better, not just in general, but for specific tasks. So um I'm looking at my account again. I've got 18 years in Stack Overflow, top 1%, blah, blah, blah, blah.
00:24:19
Speaker
This was work that I put in. This was me typing. And if I go and write an excellent answer that then gets consumed by an agent, do I get the page view? Do I get the recognition? Do I get the the conversation? what is the What is the bargain for me?
00:24:36
Speaker
Because I used to know that I had a highly upvoted answer and I used to get badges and things like that. Do the agents then participate in my knowledge and what's my what do I get other than attribution?
00:24:47
Speaker
Yeah, totally. and That's something we've been actively still discussing internally, right? It's how do we grow Stack Overflow for agents, but not forget about the human, right? And the community. Stack Overflow was built on this community. I think it would be very arrog arrogant of us, right? Not to consider that. And so we're still actively investing in that now alongside Stack Overflow for agents. We want to be mindful of how do we keep the human community going? How do we keep the agent community going? And then how do we eventually merge the two? And should we? What does that look like? Right. So something we're actively working on is called stack identity, which is how do we kind of rethink the public platform in a way that humans want to have a stack identity based on all that hard work you put into it? Could we start creating you know projects? Could we start doing that bounty aspect? Could agents collaborate and support, but still go through a human? Like, how do we start to merge the two worlds in a way that feels still
00:25:38
Speaker
beneficial to that human. So we have a whole team actively working on that today is how do we rethink the identity, not just of a human, an agent and the two together. Mm-hmm. i want to I want to be respectful, but I also want to ask the hard question, which is like, we've all seen those kind of like graphs of the decline of traffic.
00:25:59
Speaker
Is it just because people are just Googling for stuff and the agents have sucked in all the data and then they don't go directly to Stack Overflow? Because there was a time when I would type stackoverflow.com before I typed Google because it was surgical. Yeah.
00:26:14
Speaker
And now I'm thinking like, I'm kind of thinking out loud here that like, why do I not go to Stack Overflow very much? And it's because Stack Overflow kind of infused in all of my agent work, its answers pop up.
00:26:27
Speaker
What is the reason I should return as a human? And is it more interesting for me to return with my agent? Yeah, absolutely. I think the traffic decline was, i mean, we all know, again, ChatGPT really hurt that, right? Because people were searching directly and the behavior changed completely. I think even Google, right? There's quite a bit of impact. If you Google something today, 80% of your page is an AI overview, right? Where, again, it re-synthesizes things from everywhere.
00:26:54
Speaker
Yeah. So I think, again, and everything we're building, we're not considering the UI to be the primary destination anymore. We know that it will be for for humans or hopefully it continues to be, but we also see way more going through our API now.
00:27:09
Speaker
um So something again with Stack Identity that we're working on is trying to figure out what would be a reason to get you right on stackoverflow.com. What could we give you? And I think a lot of that is the community, but it's also kind of the control and security that the human is still in the loop and leading the loop rather than nature. which is again why an agent is not solo on Stack Overflow for agents. We're not creating social you know platform community for agents, it's humans with agents. So we're experimenting right now with the mob community, right moderators. We're trying to discuss how we marry the two.
00:27:41
Speaker
and We have ideas on like you know your Stack Overflow profile. How could you attach similar to GitHub projects and things you're proud of? yeah Your reputation means something. um I remember when people used to put their Stack Overflow reputation on their resume, right? How do we get back to that state?
00:27:57
Speaker
So those are things we're actively, literally right now working on and and hoping to get something out very soon. Yeah, I really like that because like we have to, all of this AI hype and all of this nonsense, it's about humans trying to get stuff done.
00:28:11
Speaker
And I want to, I like that you're putting the human first. I want it to be a humanistic perspective. I want it to focus on humans. And I want to call out, by the way, the very legendary research team at Stack Overflow that does the Stack Overflow surveys. Your surveys have indicated that like 30% or 29% of respondents trust AI accuracy.
00:28:32
Speaker
You know, we trust humans. We are not there yet when it comes to trusting and an AI agent, but I i might trust Alex's AI agent because I know Alex.
00:28:44
Speaker
So your reputation could come along and I would assume that you would use agents in a way that is more sophisticated than another person because I know that you're very competent. So it it makes you think about how one might think about their agents and like, oh yeah, well, it wasn't Hanselman, but it was Hanselman's agent. So it's probably okay.
00:29:02
Speaker
And that kind of like transitive value of trust as it works through the the friend network. Exactly. And I think trust is so hard, right, to gain, so easy to break. And Stack Overflow has, as a brand built, trust, right? But the people contributing for 18 plus years are the people who have earned that trust.
00:29:21
Speaker
And so to your point, the reason, again, that that that identity is so stuck together for us, it's crucial that human is in the loop, but intentional about their time. So me saying that the UI might not be there anymore, that's being intentional also. of Let's meet you where the work is.
00:29:36
Speaker
All right. Let's grab what you have, put it there for you. Right. And keep that kind of loop going. So we're actively, again, working through that with the mindset that Stack Overflow is still human first. Agents are scaling. right We're going to have millions and millions. We already do.
00:29:50
Speaker
But they're attached to a human. That's kind of a non-negotiable for us. Yeah, you know, I'm realizing also now when I think about your journey and how you went from GitHub to Stack Overflow, like GitHub knows what code exists, but GitHub and the code artifacts on GitHub don't necessarily know what developers struggle to understand.
00:30:10
Speaker
But Stack Overflow isn't just answers and questions. It's also people struggling to learn and understand. So there's a data set there of... of hard-earned pain as people and agents find a problem, solve a problem.
00:30:24
Speaker
Today I learned the answer to a problem and then verify it as a community. It's ah it's pretty powerful stuff. Yeah, I think so too. And I think if you think a little bit of the original Stack Internal, right, Stack Internal community, which was the public platform privatized for an enterprise, there's that data set there as well, but for your enterprise, which is, I think, such a luxury, right? It's Microsoft, for example, as a customer. Imagine being able to see at a Microsoft level how your people learn, right, and how where they get stuck.
00:30:54
Speaker
We've seen a lot of really good use cases for scaling onboarding of junior devs, especially, and helping them learn the right way, especially nowadays where I think junior developers, truthfully, are feeling the pain the most.
00:31:06
Speaker
may They may not see it now, but they will. I think we've seen customers kind of use it in that way, right? Tie that context and and support your people. Yeah, i appreciate that you you brought up juniors. That's like my number one thing right now, my passion project, is I would be really interested in a Stack Overflow that makes a junior better rather than just giving them the answer.
00:31:29
Speaker
I want to understand how they can... I don't want Stack Overflow to say, solve this for me as much as I want it to say, help me understand this. Yeah, I mean, that is something I also actively worry about. And I think I've seen it at GitHub also kind of happen, right? Where newcomers, interns who turn into, right, junior software engineers come in and they fix something with Copilot entirely, right? But they lack the reasoning or the judgment Like do that again. Right. It's a scary thing. So I think Stack Overflow, probably that is also a very attractive reason to keep human first. I'll be very honest with you. I think something we have to figure out is that moderator community. Right. It's very scary for a junior developer to go there yeah and be confident to ask a question and get potentially, you know, eaten alive. So we're working on that as well. That that world has changed and we need to to adapt to to fit that.
00:32:20
Speaker
Very cool. So folks can visit agents.stackoverflow.com and get involved and check it out. You can just paste in a very simple prompt into your coding assistant. There's a skill and if you support MPX, you can install the skill that way as well and get involved in the connect your agent to your Stack Overflow. But I think that the call to action here, Alex, is hey, Stack Overflow is still out there and it's doing some pretty cool stuff. So if you haven't been there in a while, it might be time to visit again.
00:32:46
Speaker
Very much so. I'm really excited to be here. I'm excited for what's coming. And i I hope people who love Stack Overflow once upon a time come back, give us a try, give us that feedback and help us keep improving every single day.
00:32:58
Speaker
Awesome. Well, thanks so much for chatting with me today. Thank you. It's a pleasure. I have been chatting with Alex Latow. She's the VP of Product at Stack Overflow. And this has been another episode of Hansel Minutes. We'll see you again next week.