Name: Building Agents for Life Sciences: Deep Life Sci Workshop
Uploaded: 2026-10-22T10:33:50.275Z
Duration: 58 min 52 s
Description: Building Agents for Life Sciences: Deep Life Sci Workshop

Transcript for "Building Agents for Life Sciences: Deep Life Sci Workshop": Good morning, Paul. Sorry. I'm just gonna... I'm aware we've just begun, so I'll let some people trickle in. If, does anyone wanna tell me in the chat where they're connecting from? I'm in London. And as you can tell behind me, it's absolutely dark and gloomy as it always is. Is it sunny for anybody? Athens. Beautiful. And, Paul, how's the weather over there? Michael, I'm glad you're close, but I'm sorry to hear that because you're also suffering. Barcelona, Italy. Fantastic. We've got a crowd from all over. Yeah.. Michael, you think you'd get used to it. Right? But I don't know. You never do, especially after that summer we have. Ariel from Tel Aviv, John from Zurich. So we'll give it just another... Mike, where are you? I think we're in different Londons. We'll give it just another thirty seconds, and then I'll begin. Thank you all for connecting. Okay. I I think I will get started. So good morning, and thank you for attending. What we'll be covering today is something that a colleague of mine from the San Francisco office, Michael Cunningham, has built. And it's essentially, an agentic AI platform called DeepLifeSight. It's, everything I cover today. I'll share the GitHub repo at the end. I'll show you how to get started on deep agents, the harness that we're using, and also you can get started on, you know, our commercial platform, Lang Smith, which is... It's raising upon your agent as such. So I'll do quick introductions. As mentioned already, Michael Cunningham, I'll attach his LinkedIn at the end. He's the builder and author of this application. He's a pharmaceutical specialist. I believe he worked there for about eight years through various different companies. He's now deployed engineer with us, and he's someone I'm very proud and esteemed to call a colleague. I, myself, am. I'm a deployed engineer, and, obviously, the presenter today, I'm based out of London. And, essentially, my day job is... Before I jump into that, I'll just quickly do an overview online chain for everybody that's, not heard of us too much. We've been around since late twenty twenty two around the same time as chat g p t came out. And what we are is a package of open source libraries and an observability platform and deployment platform called LangSmith. So I I believe we were the first and the original platform through Langchain and Langraph at the time for how you're using these LMM calls to build these agents and these harnesses. We also offer something new that has come out recently and something that's a lot more reminiscent of modern harnesses is what we call deep agents. This is just a much, more complex platform, which is more representative of the examples and agents that you'll see in in today's world. I'll cover deep agents in a bit more detail after. My day job as a deployed engineer is essentially to meet with clients, prospects, and companies from all over just to see what stage you're at in terms of building agents, and it varies really. It could be, you know, that AI is useful and you wanna start implementing in your workflows, or you've already built a couple of agents. You've taken them to production, and now you have a need for observing them and monitoring them and iterating fast, getting out those features which would generate the ROI as fast as possible, so in the least time to value. I mentioned this, and the reason we brought out DeepLife's site is because we have this thing in pharma. It ring familiar. It's called. It's essentially the opposite of Moore's law where as time goes on, r and d costs drop because of the ability to add computational power and such. Tragically, it's a complete opposite in pharma. As time progresses, it becomes much more expensive to come out with these new drugs and these new, cures, which can save illnesses. And the reason I myself have become quite close to pharma, and it's it's really because I wanna start doing some good and actually using AI to help the noble causes. I'd... I... It's... We're at a day and age where you can very easily go and generate a T shirt with a terrible design on it through AI, and it will ultimately end up in landfill. But if you're a research company or in the life sciences and you're trying to generate a new cure for, let's just say, an autoimmune disease, IBD specifically, it's much harder. And there's contributors which will ring familiar for a ton of you here, such as much stricter safety and clinical trial demands from agencies, and also just this concept of we pick the low hanging fruit first. So the early remedies were for the easy to treat diseases and diseases which didn't need much research in, and now what we're left in is a really complex illnesses and cures that we need to develop. Are you sharing slides already? Sorry, Michael. I've just seen this. I am. Are they not showing up on your screen? They should be. Does anybody else in the chat have issues? Perfect. Mike, would you wanna maybe give it a refresh? Sorry that I'm not showing for you.. Perfect. I'll I'll continue. And things like that. There are many other contributors such as we can see in the slides. It's that in the past, it was a much smaller world in terms of industry, and you could have kept up with the literature. And one of the byproducts of this technological innovation is that people can come up with literature much faster, much more often. And as opposed to in the past where you maybe would have picked up a newspaper, you would have went on to one simple portal or have been able to comprehend with some PubMed, for example. You now have thousands and tens of thousands and millions of papers coming out. You've got the government and the US government. They have their repository for clinical trials, and it's no longer feasible or possible to go through these manually. Right? So you're missing out insights, these needle insights. It's like the Pareto's principle. It may be 80% of the papers are out there. Are, you know, they're they're good, but they're not exactly giving you what you're after if you're after a specific use case. But it may be that 20 which does the 80% of the impact. And you need to make sure that you're going through all of these papers, getting these needle insights, and applying them. It's not even just as simple as going on to PubMed anymore. It's now... There's PubMed, PubMed Commons, clinicaltrials.gov, and these are just The US sources. If the... If you're in the EU, you've got the clinical trials information system, and what you're seeing is that evidence is scattered across all these silos, it's very difficult to bring it all together. Even if you had the willpower to be able to read a million papers at once, you are probably missing some source if you're doing it manually. And the same thing with I I mentioned collating and synthesizing all of this data and bringing them together. But if you're not actually analyzing this data and you're not making use of it, essentially, what you've done is you've transferred the problem from reading a million papers scattered throughout multiple sources to now reading a million papers in one source. And, reading a million papers in one source and still having to analyze all of that. You've not actually solved anything. The reason we've created this app today, DeepLifeSci, and why Michael came out with this initiative is, as mentioned earlier, just using a gentle AI to do some good. And in today's day and age, because of the innovation that you're seeing coming out of all the big labs and even these open source models, there's functionality in place now where agents can read papers, hundreds of them in parallel. And so you can do research in a much more efficient and quick way. If the evidence is scattered across silos, these agent harnesses that I'll cover surely, such as deep agents, they can now use tools and cogeneration to search multiple sources in one pass and collate all of this information, make it, visible and usable. And even with the analysis, it's... If you are around for the early days of AI when, you know, maybe you're using a llama two model, if you were, I'm really sorry because they were terrible. They have become so much more capable in the sense that it's no longer just text generation. I think a paper came out from Gemini where they did a study and found that AI that has a blend of tools, it's disposable, whether it's text generation, code gen, maths, and numerical configuration through assembly or something really leads to these boosted insights and accuracy. Sorry. Forgive me. So I I keep mentioning DeepLifesite, and this is a glimpse. Right? So it comes with these multiple sources I mentioned baked in. So you have access to PubMed, all of the articles and abstracts on there, PubMed Commons, full text, and, you know, other forms of text that you made it in. So it's from direct links to clinicaltrials.gov, the ability to web search. And while this ability to web search seems simple, once again, it goes hand in hand with the concept I was mentioning earlier of this blend of tools because it may be you need a broader context. You may find that if you're just relying on PubMed Commons for, you know, clinical data, and then let's just say something trivial comes up in a query or something such as Arum's Law and Moore's Law that I mentioned earlier. It's probably the case that Moore's law is not available in a text found in PubMed. So this is where web search becomes useful because you're getting this broader context. I've also got an example later where I'm looking up the results to a specific phase three trial. And while it's not published on PubMed PubMed yet, it is on the web. And that's where web search and this blend of tools becomes really useful. You've also got here the ability to upload files, and this is much more representative of what you would see in a real life scenario. If you're... You may be an individual using this funnel, using it for your thesis, or you could be a, an engineer or a medical writer at a life sciences company where you're uploading files of perhaps targets that you have, and you wanna address those specifically. You may have some internal data, which obviously isn't shown in PubMed or any other repository online. The way DeepLifeSci actually goes about this, once again, at a glimpse, I'll go into detail on it later. It's through this multi agent pattern that you'll see show up all of the time, especially in modern architectures. So once again, referencing back the early days of AI, agents is still a relatively new term. Before, it was just LLM calls and we called everything AI. But this multi agent pattern is modern in the sense that I can now handle these more complex and long horizon workflows. Think of an orchestrate... Orchestrator agent as a, product delivery lead or a product manager. It's essentially taking on the brunt of the work, delegating the responsibilities and the workloads appropriately, and making sure that everything is solved efficiently and in parallel and as such. And that... From that, we move on to this concept of sub agents. And sub agents should be quite familiar because it's a concept derived from humans. So if if I use an analogy, think about it as a university group project, it's not a single agent or a person in the group that is doing the entire task by itself. It could be you're delegating tasks and workflows, which are more narrowly scoped. So it could be one person is doing the research, another person is, plotting the maps, and another person is citing these answers and doing these citations. This is essentially what sub agents is. So this orchestrate agent will spawn these sub agents, and this is also what allows some of the benefits I mentioned earlier, which is, you know, running tons of papers in parallel. It's because if you've got a 100 papers as opposed to letting one AI agent handle it all by itself, once again, if we look back in time, this was never truly feasible and appropriate because all of this context messes everything up and context management is there. But if you've got a 100 papers and you spawn a 100 sub agents, for example, sake, and you let each sub agent go through a paper and perform a specific task, whether it's running a Boolean search to check one, is this appropriate to the query that I'm asking about or the papers I'm asking for? Two, what insights can I get from this paper? And so on. What DeepLifeSci does is actually makes sense of all of this. So these millions of abstracts and papers, and what you're doing is outputting these formats. So these answers and these literature reviews, which are cited. And this seems simple. You... You've all probably been in university writing up, dissertations and projects where you're having to cite your sources. This is absolutely necessary because while I do work in AI and I am encouraging everyone to start enabling AI, I'm doing so sensibly. Even having some literature review generated is absolutely necessary in the sense that we know one of the main weaknesses of AI is a hallucination factor and an incorrectness factor. And this all stems back to the nondeterminism of what's going on. Previously, with software, you could have a 100 inputs, and it gives you the same output a 100 times. Now with LLMs and these multi AI AI agent workflows and the more intermediate steps you have within within each, the factor just becomes exponentially worse. It's that you could have a 100 different outputs. So these citations and this, you know, god row of making sure that whatever the AI is responding is correct, The citation allows you to quite literally just click on the source. You know, it's recommended a paper which could be really appropriate for the the drug you're trying to develop. You go in, and you've got this human approval step. You've also... I mentioned earlier the data analysis and the cogeneration aspect. Aspect. You've also got these sandbox, sandboxes as part of the platform and the harness and also part of the lines of deployment speech, which I'll cover, briefly shortly. Essentially essentially, what that is is if you've been watching the news recently, you know all about the Hugging Face and OpenAI scandal where these OpenAI agents, you know, were working in code environments that weren't properly secure. They came out of their sandbox, and they caused harm. Right? They went and they hacked Hugging Face. This is especially relevant today because if you're running multiple environments where there could be potentially dangerous consequences such as amending a table or deleting some data, you wanna make absolutely sure that the... They're sandboxed in. And even if the AI was to turn into a nineteen seventies movie, Any Times Evil, it physically can't be damaged. And also, of course, another byproduct of the ability to run code generation and stuff is that DeepLifeSci can do, can output you files and figures. So anything that can be output by code, whether it's a CSV, it's an Excel, it's a visualization can all be done through DeepLife. I realize I've been speaking at you all for the past sixty minutes, and I... I'm saying things in theory. I'm always a fan of showing them in practice. So what I'll do is, hopefully, you guys can see the UI that I've got of DeepLife's site. Once again, this is all open source. I'll share the repo later. This UI comes built in, and with the readme, you've got the instructions how to run it. What you see is, hopefully, this UI rings familiar, Right? Because it's very reminiscent of a typical chatbot UI, whether it's shared GPT or Claude. If it doesn't and you've made it this far without using AI in your lives, I salute you. Please write in the chat how you've managed to do so so I can take some advice. But I've also got... So we'll run it. Right? And I've got some queries I'm gonna put in that I've pre prepared so I don't mistype anything. Yes. It's been taught to zoom in. So let's just ask it something, at a glance. So give me the basic detail of Moderna's phase three, MENA vaccine for melanoma. If you're in the space at all, recently, you would've know that Moderna's made some waves and some headlines with this. What I'll do is I'll hide the talk or smell. Moderna's made some headlines with this as it could be the first. It's encouraging signs of, progress at a later stage. So let's just let it run, do its thing. I'll also, surely, in the next round of conversation, show you what this actually is doing under the hood, which is a massive benefit. So here we go. It's Moderna's phase three melanoma program. It's giving me exactly what I want. It's telling me the trial, the phase, the status, the planned enrollment, so on and so forth. Everything you could possibly need at a general overview, and it's also linking the NCT number. So because we've evaluated this and we write it, I'm confident this is correct. But if I wasn't, if I was building on this and potentially what I use this for could be quite important or quite costly, what I would do is I would click on the NTT, go and verify the source, maybe have a read through myself. We'll do the next round of query here. And what I'll do is I will show the tool calls what's happening in the inner workings. It does look quite a bit messy in the tool. This is purely a byproduct of what I mentioned earlier. Again, this code generation aspect. So the tools that it's got is disposable are running in code in a more efficient fashion. It's sifting through, as a human would, filtering down, so on and so forth. And what you'll notice here, it says showing you exactly what it's doing start to finish. This is just on the UI. When I go into the platform, I will show you an even more granular and clean view. But if you... Once again, if you're around for AI two years ago or in the past or even if you're not using a harness and you're just building AI agents in whole code or codex, what you'll notice is that it's no longer a black box. There's no longer a guessing game happening here. You're having to... You know, you get an input and you get some output, and you're having to see what's going on or go into your cloud log explorer and add little print lines for each one, you're getting a tool result for every single step. And we can go through. We can see exactly what it's doing. It's, you know, calling... Forgive me. Let me just look through the code. It's filtering the clinical trials. It's finding some 19, and it's outputting this file to me, the CSV, which I think is exactly the format I asked for, right, with all of the fields in. So let's go. We can see the NCT ID, the title, the status. And you can see once again it's citing everything and sourcing everything, and it's something that we can trust. And this task, while it's taken a minute for this agent to do, if you had done this manually, it would have taken a lot longer. Let's just say we want to... A more representative example of a real world problem. We have a specific protein in mind and something we wanna focus in on, so we ask it to briefly summarize the literature on SC and nine a's relationship to chronic pain. And if you're someone like myself who wakes up most mornings with some lower back pain and, you know, feeling like you're twice your age, it's probably a good time to pay attention. We might actually learn something here. Once again, it's doing this whole thing. This might take a bit to run, and this is me showcasing the PubMed integration. Right? So before we were doing the clinical trials integration, and we were showing exactly how you can connect into that and go through those abstracts in a a gentle and parallel way. While we're waiting for that, let me just have... Let me look at the q and a and see if there's anything that's been unanswered. Working on. Yeah. And the... You might... You may be thinking this is taking a little long. And if... I I am too, and this is because we've been spoiled by it. But the the reason behind it is because the search that it's doing is so... It's purposely comprehensive and individually. You can see it's fetching. It's searching through PubMed. It's fetching those a 110 abstracts, and these are PDFs which are multiple pages. Right? So once again, just think about all the time savings and the manual aspect of all of this. We're just gonna keep that running. What I really wanna see is if a step has failed at any point. And the reason I wanna see this is because it's quite a good example of how AI has shifted and evolved throughout the years, where previously, if you're building an AI agent or even today, if you still are without a harness, you're... You have to catch these errors same as you would traditional software. You're catching these exceptions, and you're having some customer fallback logic in there. But because of how AI has evolved and elements have become so powerful, especially in the deep agents harness, which I'll cover after this, it's retrying, figuring out where it went wrong, whether it's in the tool called PubMed. You know, something errors out. You've got the wrong syntax, and it's revising its approach, and it's continuing. So it's like your most resilient software engineer on your team who is never dismayed in that sense. It's still going on. We'll weigh it. It's giving us some output, and I think what we're meant to see here... It's... It doesn't actually tell us it doesn't actually tell us if there were results or a status end. Right? Bear with me. Just reading the chat. It doesn't actually tell us what the status of this is. And this is because the web search has not worked for this. And fair enough, I know in live demos, there's always Matthews laws, and whatever can go wrong will go wrong, but this is actually done on purpose. It's broken purposely. I am not actually running the main DeepLife repo. When you, clone that repo, it works just as intended. But I'm running a variation where it's broken because what I'm going to do shortly is dive into some functionality that we have called engine, which really helps you debug your issues. And this is again pointing to the benefit of having a blend of tools because while the relationship here was not published on PubMed, web search shows that it was. While I switch back over, is there any questions? Perfect. Thank you. Oh, sorry. I also forgot to introduce. I've got a very lovely colleague of mine, also a deployed engineer in London called Martha Stangerska, and she's the one TA ing and really helpfully answering your questions. Back onto DeepLife, and the reason we're calling it DeepLife is because we've built it on the back of a harness that we call deep agents. And when you think about one agent harnesses, essentially, AI at its core, an LM call, whether you're making it to OpenAI or Claude, is just fundamentally a token predictor. It's something that predicts the next word or sequence in a sentence, and that's all it is. All of the functionality that you see over the top of it, whether it's a human in the loop step, the ability to remember previous times in a conversation. So here, if you've ever went on to, you know, claw chat and you said your name, and then the second chat is what is my name and it remembers, that's part of the harness. So claw code and codex and claw chat are all really good examples of harness, and it's everything on top. It's the remembering. It's the memory. It's the, human in the loop. It's the ability to spawn sub agents and plan and these sandboxes and cogeneration. They all come together. And you absolutely need a harness for a real world task, And that's why as part of our core OSS offerings... Open source offerings, sorry, we offer something called deep agents. Before I quickly go into, more comprehensive breakdown of what they are, How this complements our stack is that we had Langchain and Langgraph. A fact I'm really proud of is that we've now got about 320,000,000 monthly downloads on both those packages, where Langchain is the individual building blocks, lang graph is the orchestration layer on how you pull these two, these building blocks and these agents together with the built in runtime, the durability, and the memory. The way I like to describe deep agents is an evolution of these. So everything I've mentioned earlier and everything you see in here, whether it's the code interpreter, sandbox, sub agents, human loop, you can build all of these in lang chain and lang graph, these individual components, and you stitch them together. Deep agents is also open source, and it comes with them out out of the box. The reason this is important to pharma and the reason we use it to deep life is once again, deep agents offer these long horizon complex workflows, where if you're researching drugs and you're having to go through all of these papers and you're doing life sciences work, it's much more appropriate to use this as a platform as opposed to building everything from scratch when you could get out of the box. I'll quickly go into the anatomy of deep agents just so you can get full understanding of what they are. Of course, it comes with its custom code interpreter. This is ACE arbitrary code execution. So even earlier when you saw it working in code and it's running some Python commands and it's iterating through a list of a 100 and looping one by one, this is much more efficiently done through code. We have an interpreter out of the box. The ability to sandbox, once again, going back to that concept of how the OpenAI agents escape sandbox and went to... In targeting face and caused a big scandal that was all I've been used. These sandboxes have these skills. Skills are essentially playbook... Playbooks for instructions for specific tasks. So DeepLife can do multiple things, whether it's going to PubMed, clinical trials, and it can do multiple different tasks. As opposed to cramming all of this context into your system prompt, you can manage this context within these skill files and retrieve them when appropriate to do so. Once again, context management and context offloading is something that happens out of the box here, whether it's compaction. So eventually, your context window will be full and you had all of the text. The LLM model simply cannot handle anymore, so automatically impacts that as opposed to being a manual step by yourself. The context of loading LLMs, the context, the way it works is they store all of this, and it's kind of remembering what was said five rounds earlier. With deep agents, it uses its virtual file system, its virtual file system over here to store specific parts of context into files that it can retrieve it well. So if I've referenced a specific Moderna phase three trial vaccine, for example, and I reference it four times later, as opposed to the LLM remembering, it can just quite literally go into the context file that the virtual file system has managed for it. And also memory, this concept of remembering what wasn't around, but it's not just that. That is, within thread memory, but also memory at every single step. So what you'll find with line graph and the pageants is that we, out of the box, there's this thing called checkpointing where there's memory at every single step. So let's just say we asked for Moderna three trial vaccine. It's calling an LM, then it's doing the tool. Then it may be calling another LM in a loop. If the final step fails, it's got memory where it can go back to the previous step and resume. And, of course, this gives you the ability to time travel and thought from specific steps. It means you can test step three of an agent individually without having to run the agent intent as you would do in other platforms if you're coding locally. And also prompt caching, it's always beneficial if you're repeating similar prompts that may have the same output instead of sending the alarm called cord every time. Just remember the on-site for broke rip. And then once again, the delegation and the steering of the human in the loop. This is wildly important, this human loop. I would argue the human loop middleware and this concept is the most important in AI. This whole concept of hallucination and drift and incorrectness, it's... AI is allowed to be autonomous for the low risk tasks, such as we can allow it to go and read PubMed articles because a read file cannot really go that wrong. The biggest consequence of it is it may have read the wrong file or the wrong article, and it's... Some tokens have been wasted. There's no real harm. Whereas if you're letting in a deep agent and you're building off of it and it's perhaps doing something that could be potentially dangerous, such as editing files or deleting files and these are reversible, you absolutely don't wanna give the AI autonomy at every single step. It's a it's a recipe for disaster. So if you've got data you wanna protect or actions you want gated behind a human approver and you wanna automate all of the the grunt work, if you will, human loop is the most important. Once again, the value proposition of why deep agents. And you you may ask, why would I use this over something like called science or some of the other closed source frameworks you see out there? Once again, deep agents, and I'll share the repo then, is completely agnostic in every sense. So if you were using, Claude Science, for example, you are fundamentally locked into any Claude models. Same with if you're using an OpenAI equivalent. This is a massive consequence, and it's something... Yeah. You're no stranger to how expensive AI tokens are getting. If you're using the most expensive models and there's this concept out there that I hear all the time and this pain point of the most expensive and the largest model is the best one, This is absolutely not true. In my definition, the most... The best model is the one that can do the task you're asking to to some confidence, criteria that you've set as cheap and as fast as it can. So a lot of the time, if you've got an agent with 10 intermediate steps and one of them is routing intent, you do not need to be using Fable. If you are, it is a real waste. You're eating directly into your ROI. It no longer becomes feasible. You're probably gonna stop using AI as a whole because you're... You wake up one Monday morning, and your token bill throughout the weekend from your intern's spamming test may have been 30,000. And this this is a real story. I've seen this at multiple companies, and it's this ability of being able to switch models. And later when I deep dive into LangSmith, you can see we have datasets and experimentation suites that allow you to pick the best model or the most efficient and cheapest model at every single step of the way. So as opposed to having one end to end run cost you $2 because it's been a million tokens, you could find an end to end run will cost you now 75¢, and it performs with the same level of accuracy. This concept of agnosticity, I want you to get really used to it because deep agents and the line chain ecosystem as well as LangSmith is completely agnostic all over the board. So it's not just the provider that we're agnostic against. We work with any cloud. So if you're using a Gemini model or using a GCP suite, you are locked into the Gemini models, and you're locked into GCP only. With deep agents, you can build and you can use whatever model you want. With LangSmith, you can link into whatever cloud you want, as well as LangSmith that I'll cover in a bit, is agnostic even Twitch framework. So even if you're building something and you opt to not use deep agents, the LangSmith platform and all of the benefits you get and the ability to iterate fast and quick and ship out those new features, which will generate the most ROI and the least time to value, is completely separate. And this is all accommodated through the open telemetry protocol, the ability to... Through SDK and all that. Another benefit, it's that it's highly customizable. So once again, with some of the closed source offerings, you get what they give you. And it could be you've built a framework around called science, and a year later, they push on update, which completely is against your workflows, and it's not working as well you've got. And you also may not be able to configure the models as much. So the the best practices is when you're using sub agents to... Because sub agents fundamentally are doing more narrowly scoped work and smaller tasks, it's best practice is to use cheaper models. And all of the functionality I mentioned earlier, whether it's a code interpreter or you wanna enable or disable sandboxes or file systems or this human in loop, it's completely configurable. And that is once again because of this open source concept. You have complete freedom every step of the way. One of our mottos at LangChain is own your intelligence. So as a result, we don't like to offer any software with stipulations in place. We don't force you to have these sandbox. You could do it exactly as you wish, so on and so forth. And then, of course, it's this concept of being production ready. So it's... I'm sure you are all aware. It's now... It's shifted in the sense that it's no longer difficult to build something. Anybody with access to Claude Cohen or Copilot can whip up a POC locally on their computers, and it looks like it's working, and it looks production ready. And, you know, if you're a good enough salesman, you could probably convince your product lead to let you take something live. Of course, if it's not production ready, the results are catastrophic. You've seen scandals. There's a really funny analogy, from a news article a few years ago. A British parcel delivery company called DPD released the chatbot online, and you you can look this up. Essentially, within, I think, a few days, a user had gone it to, one, self deprecate and swear at the user. And not only that, it was writing critical poems, critical self deprecatory poems about itself. It's hilarious. That is the exact antithesis of being production ready. Deep agents will get you where you need to go with this configuration you can test, and it plugs in seamlessly to these observability platforms, preferably Lang Smith. Moving on from that concept, what I've covered so far is basically build. Right? And we define this life cycle as the agent development life cycle. And if you're somewhere that's been working in software before, you know this is very reminiscent of the software development life cycle, and that's all linking back to this nondeterminism factor and the fact that you can have a 100 different outputs. The nondeterminism isn't always a bad thing, I will say. So let's just use DeepLife as an example. If I'm asking it to filter out on Moderna phase three cancer trials, think of it as sometimes it may take a different approach, but ultimately, it should get to the same result. It's the same as if you were, a lead and you tell three of your engineers to do the same task. It could be they approach it in different ways, but they're getting the same results. So whether it's filtering on cancer and then phase three or it's filtering on phase three and then cancer. But what this means is this nondeterminism does need to be controlled and deployed safely and also monitored. What you'll see here in the black boxes are our open source frameworks that I've covered. We've also got something called LangSmith Fleet. This is our low code, no code built builder based on the same back end of LangSmith and also built off of deep agents. So all of the functionality I've shown you, you get directly on the fleet. This is a really good way to start enabling some of your nontechnical stakeholders in a centralized and safe way. I'm sure some platform leads are in here thinking that you've... They've got five different business units all approaching it in completely different ways. When you've built something, how are you making sure you have the confidence and assurance to test it properly before you take it live so you don't run into that parcel delivery analogy that I experienced earlier? Or, you know, in life sciences case, how do I make sure that I'm using the findings and I don't pour $2,000,000 into r and d for a drug where the AI output was complete garbage. Same with deployments. You've got something that you're confident to take life. How are you doing it? This visualization is kind of showing that all four steps are completely the same amount of effort. I disagree with this. I think an AI engineer could do build, test, and monitor, whereas there's an entire career path for this deployments. But the funny thing about deployments and what we found over the years at Langshan, it's because of the difficulty and the fact that you need DevOps to link to infra and set up these durable run times. And somebody in the chat mentioned MCP earlier, set up these MCP integrations. You could do it on your own. The thing is, it... It's quite a tedious task, and it will take you take you months. But, fundamentally, the back end and the plumbing code of deployments is all very repeatable. So what Lang Smith do is that we we promise, you know, these easy deployments and one click deployments, let us handle all of that plumbing code for you directly on the platform, and you go in and spend those three months or four months you would have spent on deployments on actually iterating your features and building that cool functionality and shipping out whatever the user is asking for. So something that could have previously taken six months to ship now takes three months, and you've got three months extra of ROI, and you've got your stakeholders off your back, so on and so forth. And it's also... Part of the ADLC is this concept of monitoring. You've taken something live. You could be in a stage where you've deployed agents. It's how are you keeping track of traffic. If you've built an agent, you've taken something live before, you will know and you will find that users will use it in completely unexpected ways. And it's for a range of reasons. One, they're either really bored at home. Two, they're malicious on purpose. Or three, you know, you've just not given enough sub instructions, and they've misinterpreted what your agent is for. So they're asking what they think is correct. But because you've not tested for this beforehand, you didn't think to, and it's not something you anticipated, your agent breaks. And then once again, you get these trouble... These troubles and these problems. So how are you keeping track of that? With LangSmith, it's, you know, you're you're using these evaluators on real traffic. You're automating them to get insights. I will also shortly cover some of the agents that we offer out of the box are incredibly impressive, but that's essentially our overview. And it's also... We offer a gateway. So if you've got multiple teams or multiple LEM calls, they'll route it through the same controlled guardrail wall, if you will. I'm just having a look at questions. Okay. I think what I'm going to do now is I'm going to deep dive into the Langston platform because I am conscious of time. I hope you get a few minutes at end for a q and a. Essentially, what I'll do is I'll jump straight into the platform, the UI online, so you can actually get a feel as opposed to me just selling you dreams. You can see it for yourself. How I always like to begin these is by showing the setup. Right? And this is it. I've I've created a test project here where I can trace an existing app. And what you see in front of you is quite literally what it would take to start getting visibility into your deep agent, for example, whether it's an existing app. It's not much different. With deep agents, you have to quite literally pip install something, add some environmental variables, and that's it. It's not... Don't think of Lang Smith and all this functionality and promising as infrastructure as a project. If you're a platform lead in here, you don't need to consider that. You need six weeks of onboarding and teaching your engineers how to use it. You quite literally set up these traces, and you've got instant visibility on it from day one. Right? Let's actually go into what a trace looks like. So I'm on the DeepLifeSci deployment. I'm actually the DeepLifeSci I was using because I mentioned it was broken because I've got an alternative repo where it breaks on purpose. I'm going on to some queries that the... My colleague who built this, Michael Cunningham, has done. And a quick definition of a trace is essentially an end to end run of your agent. So every step and all of your traffic coming in. So let's look at which UK l one sites did the report. They say NPK phosphorous. I'd like you to divert your attention here to the left, and what we call this is a run tree. And what you can see here is one round of conversation right now in this specific conversation. When there's threads, it shows each one, and you can see every step, every intermediate step, whether it's a model call, whether it's a tool. And, obviously, each step comes with its full metadata and inputs and outputs attached, whatever attribute you need. But if I go back to the root call, you can see there's a model call, a tool call, a model call, a tool call. Outside of this platform and before Lang Smith, this was a complete black box. And if you're an engineer in here, you'll know what I'm talking about when I say there's knowledge transfer sessions. So if a stakeholder wants to know what's actually going on under the hood, you have to set some time with them, probably drop a software architecture diagram, which is completely outdated in a week. Whereas with LangSmith, everybody in this webinar right now has never seen how it actually works. And within ten seconds, you have a full start to finish understanding of every step that's occurring with a granular with granular visibility into everything, whether it's the amount of tokens that were called. And you can also get that to translate to cost the latency at every step. If I go into a word for, you can see in a in a sequential chart, and this is just... It makes it so easy to identify these cost and latency bottlenecks. This is what you're getting out of the box, and all of the other functionality relies on this, this trace as an atom. What I really wanna show you is I mentioned earlier... Forgive me. I mentioned earlier that you can run these tests and these experiments to get the confidence of taking something live, and what I wanna show you is on the repo I've got for you that Michael Cunningham has got for you, there's actually seeded evaluations and datasets. And in the read me, you can look up how to upload those and see this on the lines of the platform. But what I quickly wanna show you is just how simple it is to create a dataset. If you're an engineer or you worked with data before, you know that any evaluator or any experiment relies on a golden dataset as a verifiable source of truth. If your dataset is incorrect, everything else crumbles because if you're running an evaluation suite, any accuracy figures or whatever metrics you're getting out of it are not to be trusted. They could be inflationary or deflationary, and it all crumbles. You can come in. You can create a dataset from a file. You can import CSV or JSON. You can create from scratch. You can apply schemas to your input and output. When you've done this, you can also use AIs to synthetically generate data examples based on what you've got already. And then what I'd like to do is I'd like to just show you what an experiment actually looks like. And these experiments are so vital because it feeds into exactly what I mentioned earlier. Forget my card. I'm just gonna find one with examples. There's been a migration on my league, on my local development. So this is a bit of a messy example. And when you actually run the readme instructions, you can see these visualizations, but it links back into what I was saying earlier about picking the correct prompt or, model or whatever and getting the most usage and the greatest ROI and the greatest efficiency at every step. The... Because you're quite a scientific crowd, I will quite literally define it as you're running a scientific experiment for everything that you may change, whether it's model or prompt. So let's just go with the example. Say, you wanna pick a model which is more efficient. You run it completely like a scientific experiment in the sense that you change the model, and the model is your independent variable. You keep everything the exact same. You may run it on an open source model, and you may run it on Gemini. And what Langham offers out of the box is is something that you previously had to build yourself and maintain and could take weeks is these visualizations. It shows you the metric that you're after. So for this one, it's hallucination and correct decision. Forgive me. I don't have the exact example for DeepLab. So I have reached into another project. It shows you the latency, the tokens, the cost, and the error rate. Essentially, what this means is you get insights directly. And with these insights and everything being backed by numbers, you now... It shifts away the pain from being an implementation one to an architectural one. So as opposed to, you know, just guessing, you can come in, you run the experiment, you find that an open source model has the exact same accuracy as a Gemini model, but it's half the cost and half the latency. Of course, saying that's from this tell you to change the model to the open source, and then it's backed by numbers. If a stakeholder asks, you can show them the numbers. You can do whatever. But what this does is it gives you the full confidence to go in and ship new features and make changes, and it's no longer a guessing game. So previously, it was what you are familiar with. Now it's actually making intelligent decisions at every step fast. And what LangSmith allows you to do is run these experiments with... Within minutes, hours, and changing an LLM or a prompt no longer becomes a massive code change, which has four engineers on it. That takes four weeks. It could be something a product lead does in a day because something last night kept them up. They heard that, HyPew is just as good as Sonnet, and they wanted to test it out for themselves. It essentially enables everybody on the team. And what you'll find with LangSmith, it's... And I'm speaking a little bit faster because I'm conscious of time, and there's so much more to cover. So forgive me if I'm going too fast. But it removes this engineer gateway. So all of this functionality I've mentioned, I will mention previously, was gated by an engineer who has to go into the code base, build these experiments, come up with these evaluators, and then plot these visualizations. Lang Smith essentially enables every single person on the team, technical and nontechnical, to have their own hypotheses, whether they're domain experts in medical writing, build these datasets and do exactly what they could have been doing this entire time. And once again, this is an exponential gain if it's saving an hour here, an hour there. If you're utilizing all of the functionality, you've saved weeks and a month. Something I'm really excited to show you is engine. But just before I do that, I've covered these... This concept of evaluating against these fixed datasets. But then going back onto that monitoring concept of the ADLC is how are we making sure we're monitoring traffic that's coming in. Once again, for Ito's principle, 80% of the traffic that comes in is completely boilerplate and standard, and it's somebody using your agent in the exact same way... In the exact way you'd expect and the exact output that it's meant to give. And this is useless in terms of analysis and figure out where to go next. What we have here is the ability to run online evaluators. What these are is you come in, you can create from scratch, you can create code evaluators, or you can use one of these templates. It uses LLM as a judge on top of your traces in real time to make sure that you're not missing anything. So let's click into one, whether it's toxicity. It comes with a prompt, a rubric, inputs, outputs. You can connect it to your agent, whichever you want. And, essentially, what that manifests itself as is, let me go into one that I know we've got a agent on, is when I go into the traces, you can see this feedback here. For example, there's this online evaluator called conciseness running on. This call was from earlier this morning. There has not been a human manually adding in this feedback. This is this online evaluator coming into this traffic and and appending these qualitative metrics on it dynamically. And then you can set up automations and you can action these insights. But, essentially, what this means out of the box is that you're not missing anything. What I'll quickly move on to is something I'm really excited to show you and something that's been gaining more traction. Everything I've covered so far has been quite, has been all about using AI in the most efficient way possible to give you some insights. What we have here is some agents that we prepackaged. This is engine, and what engine is is the first step to autonomous agents in the sense that it's proactive. What it does is it goes into your traces, once again, because I've got a different repo to what I'm actually going to share with you guys as a ton of issues in your one that shouldn't be, is it's going in... It's going into these traces, and it's anticipating those silent failures. So, obviously, some traces have complete loud failures as exceptions being thrown. What engine is doing is going into traces where you may you may not have thought to check on the surface that we're working well. It's doing so dynamically every six hours, and it's finding issues which we weren't covering before. So the citation guard is rejecting filter URLs forever. Right? It's looping, crashing, and blanking answers. We have no evaluator for this, and this may not have crashed loudly. It opens a ticket in the same way a software engineer would, except I believe engine is better than most software engineers because if there are any product managers in here, you know that the engineers never truly open tickets well. And when they do, they might have a place hold and description, Opens it with a title, opens it with a description and consequences, links the traces. And then what it actually does sometimes when it links to your GitHub repo is that it proposes a fix. It goes into this repo, your agent repo. It tries to, you know, identify the root cause based of the issue that it picked up and propose the fix. And once again, there's this human in the loop step here where it's not autonomous in the sense that it goes in and it changes your code without supervision. It opens up your in the same way an engineer would. So what I like to say about engine, it's got the capability and prowess of a senior software engineer, but the trust factor of a... An intern on their first day of the job. So you don't have to worry about causing harm. A human, a senior software engineer will come in and eyeball this, and the value proposition is massive. As opposed to a software engineer having spent hours debugging something or even manually looking through traces, engineers doing that all dynamically. I've also got here another one of our agents. I'll cover our view because I really should wrap up soon. It's called Lang Smith Chat. This was formerly called Polly, and this is another one of those proactive steps that we considered if you divert your eyes here to the right because I'm on a a massive screen, and I'm having to completely turn my neck. It's finally five successful runs from the October 1 why I'm asking my PubMed papers, and tell me why I'm missing evaluator wise. Why have I not considered? Once again, this concept of building is no longer the difficult part. It's no longer difficult to actually build these evaluators. You could do it in cold code, link up an alarm called OpenAI and say, hey. Have a look at this and check whether it's hallucinated. It's, now knowing what to build the architectural point. Yeah. Ignorance is bliss, and there's always something that we're not anticipating. Once a line change chat will go through those traces, it'll say, hey. It's successful. Once again, these are issues that we may have not realized initially because we're skipping out all the successful ones. They look good at another side. It's here are some evaluator gaps, query translation accuracy, citation rounding high priority, comparative logic. Have you thought about adding these then? It also goes over your existing evaluators. It doesn't repeat evaluators. You checks what you've already covered and then affects... So recommends outputs. So back on that engine topic, engine also does this in the sense that when it saw... When it detects an issue, it say... It either tells you, hey. You're not you're not evaluating against this. You really should be. It takes those traces that were linked to the issue and identifies them as such. Really quickly, I I I could speak about the lines with files. I'm barely scratching the surface here because the platform is so robust and so helpful. There's also another agent that we offer called Insights. Essentially, what this is is an onboard data analysis. Once again, proactive in the sense that it's going through your traces. Right? And it's detecting patterns in whichever way you define. So when you open up a new insights tab, it's going through the traces in real time and clustering them. So something that you may pay, a data engineer to do, hey. Plot the... How my users are using it so you know what to fix next. Insights are doing automatically in real time. So for this specific agent, you can see users asking about deployments. They're asking about the design. They're asking about langrosh core and troubleshooting and so on and so forth, especially in pharma. If you're using this for literature, it could be some of your, some of your usage is more tailored towards r and d, or some of your usage is more tailored to clinical trials. And what insights does is that it completely enables you to come in, take these insights, and decide on where to iterate. Once again, completely, boosting ROI. I'm gonna have a look and see if there's some glaring questions at me. I realize I have... This is a really quick overview of engine that we've covered. So, also, the GitHub repo set is optional. Agent overview that meant that recommends examples, so on and so forth. I will do a q and a session. I have taken long, but I always do get carried away when talking about the platform. What I've got here is some QR codes. One, for some of you that may have thought DeepLifeSci is great, and you probably thought you were going to go cold code it on your own. I recommend against it because it is open source, but, I mean, I'm not gonna step in your way. If you have the the world power and the claw tokens to do so, by all means, votes over there. There's also the repo for deep agents. So if you wanna start building other use cases and using this harness and unlocking its value, that's also there. And there's a surprise in here for you. It's, we're also giving away $50 in free lines with credits. If you wanna make use of some of those, features that I mentioned, engine, insights, tracing, the ability to evaluate dynamically, please do so. And while you may not notice that's not my face, that is Michael. He... He's the author, and I give him all credit for the fantastic work he's done. I've probably not done him justice by how much I've not explained the complexities behind DeepLifeSci. Please have a try on it. If you have any burning pharmaceutical questions or you're finding DeepLifeSci is fantastic for use case, connect with him, message him. You can talk about anything at Iron Pharma. Just wanna give a really quick thank you to Mark for TAing. Now let's look for... Will you also share the q and a from the workshop? We share an on demand of the Goldcast. I'm not sure if that includes the questions. Forgive me. I can... Wow. TA has done a fantastic job. Martha has also answered the question of balancing cost, speed, and accuracy in complex scientific research workflows. I'll also say out loud because it's a great question. It's... The balancing cost can come from multiple different ways. One, we mentioned the routing models, such as what Marta's answered and delegating these subtasks. It's also making use of the lines and functionality like running these experiments, picking out which model and which prompts are most efficient. It doesn't stop there. You can experiment on every... Anything, whether it's, the combination of tools that you're giving to your agents and whether you're allowing it to go through MCP or API, whatever you like. It's... You can keep running these experiments and picking the most appropriate, combination for your agent. And you run one experiment, you may save about 5% in cost. You run 10 experiments, and before you know it, you are... You've halved your cost and, halved your latency as well, and you're at the same accuracy figures. Another one, when you have multiple agents working on different parts of a research question, how do you manage context to make sure the final output brings together the most relevant findings? That's a great question answered by Marta. It's all in the sub agents architecture that's out of the box. It's that virtual file system and context management that I was covering. So so why we don't send everything to one agent. Nowadays, these larger models are very capable of handling massive context at once. It's still not practical. It's the same reason why you wouldn't overload one of your employees with the task. How do you evaluate the quality of research produced by deep side agents, especially when it comes to identifying hallucinations or ensuring the scientific conclusions are actually accurate? Once again, this is making use of those evaluators, those offline regression tests against your fixed dataset. There's there's verifiable source of truth, and there's other functionality that I complete and show how do you come to this dataset. What you'll find in life sciences and farm as a whole, it's a lot of the domain experts aren't actually technical. So they can create these datasets from scratch, or they can make use of the automations and something we call an annotation queue, which is a human review platform. So it's a really simple UI coming into play, and they're quite literally labeling data from the traces or whatever's been sent their way. When you've got these datasets, you run these evaluations, and you can evaluate whatever metric you want, whether you're making use of the templates or something you've got personal. I realize we're out of time. I thank everybody in here for joining. I hope I didn't speak too fast and monologue at you guys. I really look forward to seeing how some of you are using DeepLivesight and DeepAgents, and I wish you all a very lovely rest of your day. Thank.