I don't know about you guys, but this is the coolest tech conference I've ever been to. And I have had nothing to do with it. So before we get started, I do want to shout out everyone who's working behind the scenes, in particular Jess, Julia, Jacob, and Brianna for making this event fantastic. So let's give them a big round of applause.
Now yesterday, we talked a lot about what building agents today looks like and how a lot of the most successful teams are bringing agents into production. And so for today, I want to talk a little bit about the future. One of the things that we think a lot about at LinkedIn—it's actually in our mission and vision statement—is what do the agents of the future look like? And so for the rest of this talk, I want to pretend like we're a year in the future. It's Interrupt 2027. And what are the topics that we're going to be talking about? What are the topics we're going to be discussing? And I'm going to be wrong in a lot of these. This is going to look really silly and really stupid. But hopefully this gives you a glimpse into where we think that the space in the industry is going.
One thing that I think will start to happen is there will be a divergence in the types of agents that people are building. So I think there will be two types of agents, and we're already starting to see some of this emerge. So one of the types I think will be these long horizon style agents. So these are agents that run for minutes and hours and maybe even in the future days. They do code execution. They do planning. They use subagents. They maybe use multi-agent systems. They use skills. And they operate over longer and longer time horizons. Things like outcomes and goals are starting to become ways to increase the horizons at which these can operate. And so I think we'll see a big push in these as they do more and more valuable knowledge work.
At the same time, there's a completely different set of agents that latency is a huge factor for. And these often end up looking like customer experience agents. So this could be customer support. This could be sales. This is places when you're talking to an end user. Brand is really important here. Voice becomes a really interesting modality. Maybe video will in the future as well. And so I think these are these two directions that agents are going in. There is a shared stack underneath. There are also differences in terms of the technology. And I think one of the big questions that we'll be thinking about over the next year is how common is this shared stack versus how particular are the technology pieces that you need for each?
Speaking of voice, I think we'll be talking a lot more about voice in a year. This is a place where we're starting to invest more and more. We see this be particularly relevant for these customer experience style agents. Voice is a great modality. The typical voice pipeline today looks something like this. There's this speech to text, text to speech kind of like sandwich. So the user speaks, that's transcribed into text, that's passed to an agent. The agent operates in kind of like the text space, outputs some text, and then that text is transformed back into speech, which is the audio that you hear from the agent as you're talking to it. And so this is the pipeline that we see today.
But there are these native speech-to-speech native voice models that are coming out. So OpenAI released V2 of theirs, I think, two weeks ago. And while I would say the consensus today seems to be for applications where you really care about having control, they're not quite steerable enough yet, we do expect that to change. And so I think one of the big things that we're looking at in the future for voice agents is which will it be? Will it be this pipeline approach? Will it be this native voice-to-voice approach? Will it be a combination? What are the pros? What are the cons? I think we'll be talking a lot more about voice in the next year or so.
We think all agents will need a sandbox, especially these long horizon style agents. So coding is really good for a variety of tasks. It's not just writing software, it's for data analysis, it's for web browsing, it's for image gen, for deep research. I was talking to someone yesterday who said that when they think about building agents for their marketing team, one of the ways that they think about it is they think about giving that marketing team a software engineer. What would that software engineer build? What apps would it build to make the marketing team's job easier? That's what giving an agent the ability to write and execute code is. That's why we think sandboxes are really important, and that's why we launched sandboxes yesterday. So I actually think agents are very early on. Sandboxes are very, very early on. It's still just the dawn of those. I think we'll be talking a lot about them over the next year.
Open models are a place that we expect to see significant increase in over the next year. So the performance of these base open models without any type of post training for particular tasks is already approaching that of frontier models. So this is some benchmarking that we did on deep agents comparing closed frontier models to open models. And you can see that it does lag in some places, but it's starting to get really, really close to that frontier. The other big thing that is happening is cost is starting to become more and more of an issue. So there are multiple cases where, in particular for coding agents, these agents are burning through a ton of tokens really, really fast. I think open models offer a cheap alternative to these, and we expect that to be a big driver of open models.
A third reason that we think open models are really interesting is because they can be trained for your particular domain. And so I think that's yet another reason. As companies build up a lot of these traces, a lot of these agent runs, how can you use those? How can you use those to improve the model over time? I'll talk a little bit more about that in a second. But this type of post-training is something that you can do with open models. And so we expect a big rise in interest in open models.
Agent identity is really, really interesting to us. So as agents start to do more work in the real world, they take actions. How do they take those actions? On whose behalf do they take those actions? I think it's still early here. We see two emerging trends. One is when the agents act on behalf of individual users. So if I'm accessing an agent and that agent has access to Slack and I tell it to look up something in Slack, it will use my credentials and it'll have access to whatever I see in Slack. And so if Julia or a colleague uses it, they might get a different answer because it might see different things.
Another type of auth is when they have a fixed set of credentials, usually a service account or something like that. And then whoever is interacting with that agent, they will always use that fixed set of credentials. So they'll see the same responses. This, I think, started to get really, really popular with things like OpenAI, where you had this concept of an agent that was its own thing, and it had its own fixed set of credentials, and you'd expose it through separate channels, and you'd interact with it. And we started to see actually some SaaS providers make it really easy for agents to create their own accounts so that they could have their own fixed set of credentials. And so I think this will be an interesting trend to watch. I actually think we'll see both of these in the future. I think we'll have agents that act on behalf of users, but also agents with their own fixed set of credentials. And I think being really, really precise about when to use which one and which one is which, and making that clear to users will be important.
Continual learning is maybe one of the areas that we're most excited about as a company. So when we think about continual learning, we think about improving the agentic system over time. And there's three layers to this agentic system, and all of them can be improved. There's the model layer, there's the harness layer, and there's the context layer. So the model, I mean things like Sonnet, GLM4, GPT-4. Harness, this is the code surrounding the model that connects it to the environment. This is deep agents, Claude Code, pi. And then context might be what we provide to the harness as ways to guide it on particular tasks. So agent.md, skills. You can't edit Claude Code directly, but you can give it skills and give it an agent.md that tunes it for your particular task. And so these are three different layers of the agentic system, and we think all of them are involved in continual learning.
So maybe starting at the model layer first, this is taken from research done by Ramp and Prime Intellect that came out last week, where they fine-tuned a model to be really good at Ramp sheets. And you can see here that the latency is very, very low and the accuracy is very, very high. And this is an advantage of using an open source model, Qwen 3.5, and fine-tuning it for their particular domain. And so this is an example of continual learning at the model level.
You can also learn at the harness level. So there's a great paper out of MIT and Stanford called MetaHarness, where they used an agent to optimize a coding harness. They optimized it on Terminal Bench 2. And you can see that it outperforms human-written harnesses. And so this is an example where they used the feedback from the environment, the feedback from running it on Terminal Bench. They took that feedback, they passed it to an agentic system itself that would make edits to the harness. So it wasn't changing the model at all, it was just editing the harness. And they got a really good score there.
My background's in classical machine learning, and there are a lot of similarities between how you use data there and how you use it in this new world of agent development. And so in classical machine learning, you have the model, you have the training data, you do some gradient descent, and that updates the weights of the model. When you're updating the agent in general, depending on the layer you're at, if you're working at the harness or context layer, it's not exactly gradient descent, but the evals that you write act as a forcing function. So in the example of the MetaHarness that I just gave, you have these evaluations, Terminal Bench 2, you run the agent on Terminal Bench, you get some feedback, you then pass that into the agentic system and it updates it. So those evals are providing a similar type of this training gradient. And so evals and traces are incredibly important for this learning.
We did this ourselves, and so this is a few months old at this point, and so all of these, you know, the leaderboard no longer looks like this because the space moves really fast, but we moved from top 30 on Terminal Bench 2 to top five just by changing the harness itself. So no changes to the model, only changes to the harness, and we saw a big increase in performance. And so we think that more and more companies will be doing this type of continual learning, whether it's at the model layer, the harness layer, or the context layer, for their particular use cases.
We want to help people do this. And so one of the things that we're announcing today is LangChain Labs, which will be a research group inside of LangChain aimed in particular at continual learning. We think that, yeah. (audience applauding) We think that LangSmith, with all these traces already in there and all the feedback associated with it, this is a really, really solid foundation for doing this type of continual learning, whether it's at the model layer or the harness layer or the context layer. And so we're excited to work with a bunch of our customers going forward in this direction.
The last thing that I'll say is we think that everyone will help build agents in the future. And we already see this today. So we already see that everyone in an org is involved in improving the agent with feedback. So you've got UX research, there's domain specialists, customer support folks, product people, engineers, they're all giving feedback in some form, whether direct feedback or by tweaking the prompts or by just debugging in Slack, they're all giving feedback that improves these agents. And oftentimes it's the domain experts themselves who have the best type of feedback and know how the agent should perform best. And so we think in the future, they'll be even more involved in creating these agents. We think they're actually going to be building their agents. They're not just going to be giving feedback on them and handing that off to a separate team, they're actually going to be involved in building these agents.
We've started to see a glimpse of this at LangChain, where we have a bunch of agents created by people in all different domains and expertise. And we've seen them build these over the past few months. And so to talk more about what that future looks like, I want to welcome onto the stage Caroline Di Vittorio. (audience applauding)
