LangChain podcast mentions

Quote

So that uh Yeah, it was was great. Okay, awesome. So we can come back to a workflow anytime you want. Uh But I just wanna keep keep moving on on like all the stuff you had announced. Uh we should probably also just touch on the ISD. I know you're not like as closely involved to that team, but obviously one of the most successful opening source projects. I mean obviously Vercel is very good at frameworks, but I think it was not a given that AISDK would be a winner because of Langchain, because of Mastra, because of one you know Except that you guys have the perfect package name. I actually don't I'm not sure how well how much that helps, but it's Great like there's a there's a fun background from what people thought AI was for ten years. But uh yeah, I think we you know we we announced version six beta. And I think the the big I mean it's not really news because you know these things are open source and you can follow Experimental in ASDK five is a direct agent abstraction. What I do want to mention is actually'cause you mentioned it's very successful, which is true and I think The the the reason why it's successful is because we could st

Quote

only three that have really come to any prominence this year are us Google and OpenAI and we're we're Probably the biggest differentiator is the there is a there is the this a this modern definition of an Agent, uh I say modern, as in this year, maybe very end of last year, that that we Google A D K Arno uh OpenAI agents all agree around, and then there is Uh Langchain in particular who have chosen to disagree with that model of agent and they have uh Langchain at itself though like aren't the low level library for unifying requests. And then they have a Landgraph, which is their attempt at building this like graph library for building more complex applications. And so I I think yes there's people like Curry I and Lama Index, but I think the the d you know the really the frameworks from the the model providers and us. And I would say that it was a reasonably it was getting towards being a no brainer to use us, but I think um we would we would expect to overtake Landras. I think graphs uh we look we have graphs support in Python KI. low level support, our graphs are are type safe. They don't allow parallel node execution because doing that is almost impossible in a type safe way. We think durable execution with something like temporal is a much better solution. Yeah.

Quote

Harness engineering is the subset of context engineering which primarily involves leveraging harness configuration points to carefully manage the context window of coding agents. It answers how do we give our coding agents new capabilities? How do we teach it things about our code base that aren't in the training data? How do we increase task success rates beyond magic prompts? And one of the things that they point out is that harnesses aren't just one thing. To some extent, harnesses work backwards from what models can't do natively to create some component to solve for that. In another post from Viv from Langchain called The Anatomy of an Agent Harness, Viv added a chart that showed the desired agent behavior versus what the agent adds. For example, the simple one That's a part of every Claude code session. If the desired agent behavior is to write and execute code, the harness adds bash and code execution. If the desired agent behavior is safe execution and default tools, tooling, the harness adds sandboxed environments and tooling. If the desired agent behavior is remembering and accessing new knowledge, the harness is going to need to provide memory files, web search, and MCPs. And importantly, when you've heard about all of these techniques like Carpathy's Auto Research or the Ralph Wiggum Loops, those are harness additions to get to the desired Agent behavior of complet…

Quote

the just in time authorization behavior that we wanted because most Agent frameworks build that stage into the point after the agent has almost completed Whereas what we needed was the ability to really get in there into kind of like the internal gubbins of the system and it just we just could not find a way. We tried really hard. hard to make that work. So that was kind of why that built up. And then we do use both. both Langchain and Langsmith under the hood from an eval perspective, we consider ourselves a kind of a different level of abstraction. Sometimes people are like, Well, why don't I just use Langsmith or Langchain? And we're like, no No, you go ahead, that's absolutely fine. But like if you want to build up complicated multi step agent workflows you have to put in a lot of work on both the eval side and on the agent side to make that work. So that's kind of our our rationale. the biggest learning you have so far building open source projects this way because I think giving you or your co-founders background working on Stripe. I don't think this is your normal product building. experience. What has been the biggest learning you have to do or surprises maybe? Maybe even like to do this well? Yeah, I would say the biggest learning is in most startups you build you know with a bunch of like sticky tape and like glue and you just build as fast as you can And when you're…

Quote

Build the future of multi-agent software with agency, AGNT. The agency is an open source collective building the Internet of Windows. agents. It's a collaboration layer where AI agents can discover, connect, and work across frameworks. For developers, this means standardized agent discovery tools. Seamless protocols for interagent communication, and modular components to compose and scale. multi-agent workflows. Join Crew AI, Langchain, Llama Index, Browser base, Cisco, and dozens more. The agency is dropping code, specs, and and services all with no strings attached. Build with other engineers who care About high-quality multi-agent software. Visit agency.org and add your software. Support. That's AGNTCY. It is an interesting time for business. Tariff and Trade policies are dynamic, supply chains squeezed, and cash flow tighter than. ever. If your business can't adapt in real time, you are in a world of hurt. You need total visibility, from global shipments to tariff impacts to real-time cash flow and the That's NetSuite by Oracle, your AI powered business management suite, trusted by over. 42,000 businesses. NetSuite is the number one cloud ERP for many reasons.

Quote

in the space of actually providing some tools, some open source tools that some of our So one of them is agent spec, which is this solution to actually making our agents portable. across different frameworks. Because one thing we saw was this sprawl of people building their agent uh agents using different frameworks like crew Using Langchain or using uh Lama Index or A stack and we thought to ourselves we need to standardize this and consolidate it because an enterprise we actually have to have a common way of talking between large organizations and the team put together a open source tool called Agent Spec and behind that is another framework called Waveflow. So we are definitely trying to meet developers where they are either by giving the Oracle AI database. You obviously we've talked about the deep learning You also run live training. And O'Reilly, I think I might have mentioned that early in the episode, but can't remember for sure. So people who have O'Reilly subscription.

Quote

This sounds like an architectural necessity for enterprise Gen AI. It is. Instead of relying solely on the vast general knowledge baked into the LLM, RA instructs the LLM to generate answers based only on relevant data retrieved from internal, secure, and up to date sources. This ensures the output is contextually accurate to the enterprise. And the RJ architecture is complex. And third, a framework like Langchain, which acts as the intermediary, instructing the LLM to use those vectors embedded. Critically within the SAP ecosystem, HN Cloud offers the native vector engine capability. and SAP AI core provides the necessary support for the framework integration. And it's also significantly more cost effective than fine-tuning a massive LLM on all your corporate data. Which brings us to the crucial discussion on model evaluation and cost optimization. AI applications are inherently concerned. Consumption based. Evaluating LLMs involves non-traditional parameters. EAs must track the model provider, the size of the context window, speed latency, price, and the value of the process. And operating costs depend entirely on token exchange rates. The calculation is precise and has to be planned for. It requires tokenization.

Quote

necessarily obviously to to us maybe. But basically I think that's actually a big problem. When I talk to people at KubeCon, one of the biggest takeaways for me was people don't know how to run these things. Like they they build something on their laptop, they have a little bit of Python code or some TypeScript code. And they're like, but like where what do you know? It's like, okay, I have this thing, but I don't actually know where to put it. I don't know how to run it. Is it just an app? Like what's actually happening here? And I think that the typical answer if you go the like the Langchain crew AI route is that it's SAS, right? Which will work for some people, but it won't it usually won't work for these larger enterprises, especially the ones that want to deploy on on-prem. So what what we said is let's take let's actually take what we learned about building agents and simplifying them and And then wrap it in such a way that you can run and build just like you would any other Kubernetes app. And that's how we ended up at Cage. So one way to simplify that is to think about these agents as just another app that's running in in whatever environment you run your applications in. If you start playing around with any MCP tooling, I've been playing a lot with the Docker tooling because it's so easy to use and it's container native, which is like my whole thing. So that that fit…

Quote

The uh like in malicious behavior, like in intentionally deploying M C P servers are bad. for you. Now again for the other side there's not yeah there's nothing that's actually going after the They're doing it internally, maybe, but that's an that are you w are you willing as a as a product to take that risk and just trust that they're doing their Right,'cause if you think about I think it's Langchain, they have a set of tools. tools you can actually incorporate into your agent we if you build it with Langchain or L Or Landgraph, I don't remember which one is it. But they have a subset of tools that they have on their website. That you can just go and and hook in who's doing the due diligence there? This looks like an official thing from the U.S. MongoDB and one of the commands you can execute execute through the MCP is drop days out layer base. But drop collection, delete medi. In that case, what is the protection? I mean it's not in the MCP server, it might be official, but from the point of the provider, they want to give you all the functionality to this MCP server so they don't care. Yeah. So are you gonna be The you know detecting these and filtering this on the LLM level itself? I don't think so because That's what the jailbreaks and attacks that come in. That's right. You need uh normal security controls.

Quote

model that it will be more expensive and a little a little more slow. Now um I want to talk a little bit about your your architecture. I think you sort of visibly took um lang chain kind of out of your out of your stack. Can you talk about why you why you did that? Yeah, sure. There's a few different things, right? I think uh on the early days when we started Career AI, I think I think using Langchain was a good a good way for us to get access to a bunch of tools. from the get-go. And because those tools would help a lot of those people kind of like take some of those actions. I think that was kinda like all right let's let's do length here because that will help us with some of that Now as we start to grow and we start to build more and more logical things in sync and some of the decisions were starting to diverge and how like we were taking so decisions on the framework and like how they were like changing some of the code on their side. So It got to a point in time where we were overriding like ninety percent of the code. code that we were importing and that it was one specific piece class that we were importing. But we are overriding so much of that that every time that we need to update a version, we need to like go to crazy rebases. So I was like, all right, this this is not a fly. Like we need

Quote

Like, huh, I wonder if it's not working because this problem is too hard. hard or if it and it's not capable or it just like got stuck on one of the loops and it wasn't able to complete it. So I think there's a lot of that investigation work too that becomes a little bit of a nuance and and Yeah, it's it's certainly if you've if you're using uh one of the popular frameworks like Langchain, you have um limited So in in Langchap Um they s they support these build languages the framework where you allow you to create these graphs. But the graphs mutate after each tool call. They can accumulate state and accumulate see it. Obviously you can you can get a kind of a commercial um um Observability framework in place and you could have a page. But sometimes your agents will have hundreds of actions. So scrolling through what a thread of hundreds of actions. is quite limiting even for very technical users. So I think breaking the your questions into what am I trying to look at is is important.

Quote

serve the context for your agent. So it doesn't replace the agetic frameworks, right? be thought as something that can enhance your like agent framework that you use, like Pythonic AI or like Lang Graph or anything else. Built the state management, the context management using Fenic. talk about agentic use cases is the role that it plays in conjunction with some of these agent frameworks such as Langchain, Pydantic AI And the list is far longer than I am aware of at this point. Yeah, a hundred percent. I think there is a difference. I mean, the agentic framework is and it should be all about let's say like the the agent in clue, right? Like what gives the primitives to define like let's say the the broad like behavior of like your agent. Keeping in mind that the agents and LLMs in general right it's like a stateless thing. It's like you put tokens in there. Like you can't offload the state management to the agent. itself. Like it's not going like to work. So if you make this clear in your mind as an engineer that okay, my agent is let's say like the logic that runs there and it's stateless. Like anyone who's coming like from functional programming like I think it would it would feel like pretty natural, right? Like it's a function, gets like some token chain and spits up