The Missing Layer After Launch - Raphael Kalandadze, Wandero AI

All right, so we built an agent, you launched it, everything works pretty well in the demo, everyone is happy. But now let me ask you a few simple questions. So how do you know if it's actually working out there? How do you watch across hundreds or thousands of real conversations every day? How do you feel or understand the health of the system? How do you make it better? How do you find the holes that you don't know are there yet? And that's the thing, right? So most of the talks about the agents and the moment when you ship. So we build it, it works the end. But I think the shipping is the moment when the real work begins. And somehow only a few people are talking about that and I'm calling it the missing layer. So let's dive into it. So that's the world that we are living. So you can create the whole product, you can create a whole startup in a couple of days, in a couple of weeks. You can write hundreds of thousands of lines of code. You can spend a lot of tokens. And to be honest, it's the easiest part today with the help of the latest models. But I think the shipping is the moment when the real work begins because you need to close the loop as soon as possible. So after your launch, you need to have some control and understanding of the system. And from my experience, the loop is at least as important as the product itself, sometimes even more, because the tight feedback is the one that helps you to make the product better every single day. And that's the missing layer, and that's what the rest of this talk is all about. So what happens after your launch? And this is not something surprising. We had the same questions in the classical old software. You need to monitor what is happening. You need to understand how it behaves. You need to have some logs and detect the problems and fix them. And for agentic systems, each one of those are even harder. And sometimes they turn into something genuinely new. So let's talk about why this is hard and why this is hard now. So the agent is on a normal software, right? You don't have a few features, several buttons. You don't have a predefined flow that you can test before you go to the live. And the coverage is endless. So think about like plot code or codex. They can do a giant range of stuff wherever the user needs. And most of the agents do the same, right? So you give the instructions and they can handle it. And you cannot write all the conversations in advance. So this leads to the deepest part of the problem. The part that keeps me up all night, which is you lose the feel for your own system. So after you build the product, you need to have some kind of understanding, does it get better or worse? So you need to monitor, understand what is happening. And the problem is that the normal safety nets don't save you here. And Bill and Me, we try a few stuff. We build a unit test. We have some reg acts and rule based checks. We even create some scripts to simulate the customer conversation. And yeah, it helps in some ways. But at the end of the day, it is like only one slice of the whole problem because customers always do something different. You cannot write it all down. They are too many and they are all different. And Harrison put it in the same way. So you don't know what your agent will do. until it is in the production. So why this is a new problem? So as you know, LMs are not deterministic. The same input can have a different pass, even slight modification of the input can cause a different trajectory and the coverage is endless. You cannot list it all down, you cannot pre-test until it goes into production. The second and the scariest one is the failure height itself. So let's think about what happens when the agent is running for a long time. It struggled in the middle of a task. It had some problems, but it was lucky. It was recovered. It finds some workarounds, tries some other tool calls or whatever, and you did not get any red alerts, any problems on the dashboard. Everything looks fine. But you know, this is an early warning for you. This is a problem that is hidden and that lives in your code base. And you need to fix it as soon as possible because if you're talking about reliable agents, each will depend on their luck, right? And also as an anthropic method in the blog post, sometimes the agent loves to make market features as complete without checking if they actually worked. The next one is the tool calling, right? So the tool surface is huge. So long-running task needs a hundred of tools, even thousands of tools. Sometimes it has several summarizations in the middle, it uses some subagent, writes a lot of code, it uses a terminal for sure, it calls some other company services, third-party libraries, and they work for a different way all the time, so actually you don't know what you're looking for. So sometimes finish doesn't mean it is helpful for the user, maybe there was any problem, so agent finished, everything looks good. the answer was successful. But what happens, like for example, for our use case, sometimes a user asks to build the itinerary, agent runs the flow, it builds a trip, but it booked a different service, it made a lot of mistakes in calculating the price, so user is not happy. So technically it's successful, but still failing the task. And as I mentioned, the unit tests don't save you here, and production is the place. when you learn what you need to test in the first place. All right, so what happens after you launch the product, right? So you need to monitor, you need to understand and improve the system. And what's the source of the truth, right? This is a log. Everyone has the logs. Everyone loves the logs. You have structured information. And, you know, those machines are the best to understand and explore the logs, much better than any human, much faster. they can write some scripts to filter the giant of walls. And this is the most obvious way that you can hand it to the agent. But as soon as you start working on that, maybe you build an agent or skill or whatever, you quickly understand this is not as easy as you imagine, because it needs a lot of reasoning, you need to understand the problem to differentiate if it's a real bug or a noise. to deep dive in a trace or trajectory, understand if it's a symptom or a root cause. So actually, you'll find out that the operating agent itself is an agent's problems. So this is a loop like end-to-end. So you have traces, you have trajectories, you give the agent to access the core base, it diagnose the problem, understand what's happening and send the PR. And then you can have a scale or sub-agent or wherever that controls the PR. because you know, most of the agents are pretty eager to send the PR. They love to fix the problems. We prefer to have a separate agent, which has like fresh context. It tries to check the PR on a different angle, on different view, run the focus test. And it is not biased of the problem itself. It tries to criticize, score the PR, and sometimes it requires the changes. Maybe it closed the PR directly and helped us to filter those problems. And then you have the human in the loop. Sometimes you don't. And we can talk about this later. But, you know, this is the most obvious thing that you can hand it to the agent, which seems like pretty obvious for most of the people. And a lot of teams use the same practice. But I think people don't appreciate how important it is. and you need to spend some time on that, you need to calibrate, you need to make it reliable, to trust the loop, and this is the fastest loop that we've ever had. So, you know, after you build this simple system, you already have the feel and understanding how it behaves, what is happening, you detect the problems, some local fixes as soon as possible, and you can have the PR ready in half an hour. and you can easily understand what is happening. So let me walk you through how I'm handling this. Maybe this isn't optimal, but I think this will help you to get some point. So actually we have two main flaws. The first one is the fastest loop that helps you to detect and fix the problems as soon as possible. And another one is like more in a zoom out that helps you to have a high level understanding and it helps you to have hand on the pulse. So the first one is a log monitoring agent, the one that I already mentioned. So you have projectories, you have logs, you have the access to a code base, and it runs every hour or every few 15 minutes. And it tries to understand the problem, it deep ties in the logs, understand does the user end up stuck and send the PR or sometimes send a Slack alert. And this is pretty important because sometimes the problem is so critical, so we need to fix them as soon as possible. And yeah, it works pretty well. So this is the one example how the PR can look like. So you have a... nice description, a short explanation of what is happening. You have some metadata, you have nice diagrams, mermaid or ASCII tables, maybe you will have some HTML artifacts that helps you to give you a glance of the problem and quickly understand what is this PR for. So this is an example of how a stack notification can look like. It helps you to quickly give you feedback. detect if there are some critical problems. Sometimes it is just like heads up, you know, there are some warnings, there are some problems, so you need to check them when you have some time. And as I mentioned, the review agent is pretty critical because it tries to check the problem in a different angle. It always tries to criticize the problem, score the problem, and as soon as it's ready, it sends the PR to the human. And sometimes people talk about that if we need to have a human in the loop, because you know, still human in the bottleneck in this case. In our case, the PR agent and the review agent send 10 times more PR than three of us every day. So we need to have some clean system how we're going to handle this to not be a bottleneck in this system. But from my experience, as I mentioned, this PR and the review agent helps me to have a nice description, some artifacts that helps me to quickly understand the problem. And maybe I spend a few minutes to at least understand what is happening. So I think at this time it's okay, but maybe we will remove it in the future. And yeah, people are talking about that. So you need to, you don't need to be a person like in this problem. Some of them prefer. to remove the humans in this loop. But you know the trend is that you need to close the loop first. So let's make the problem when you are the bottleneck and then you can remove yourself pretty easily, I think. So this is some examples how the review agent feedback can look like. As you know, it requires some changes, you have some summary, you have some diagrams, explanation, some risks, some edge cases. And sometimes after a few iterations, they go in a loop where the key is ready. So we don't need to have any more changes. So after that, the human will jump into the loop. For the previous agent, as I mentioned, that is especially good when you want to fix some local problems. For our case, it runs every hour. It checks only a one hour window of the logs. It doesn't have any high level understanding of the problem. So we found out that we need to have another system that helps us to get more than a high level explanation, some kind of a health of the system. And you know, visibility is the easiest piece. Before the agent, it was impossible, so you cannot deep dive or summarize hundreds of conversations. But right now we can have a system that helps you to score every conversation, understand what's happening and give you a high level zoom out of the system. So this is the session analyzer and the main goal is to just give me the score of the health of the system, right? So we try to check every conversation, run through a lot of site patients, expand a lot of tokens. But yeah, it helps us to detect some patterns connecting the dots and try to understand some high-level patterns. And yeah, it also detects some classical problems, what is the cause, how many tool costs are used, how many sub-agents, how many summaries happen, or whatever. But also it gives some AI insights, some entities, so we have an understanding of the problems. All right. As I mentioned, one of the main problem is that you lose the control after you launch the production. So you need to have a system that helps you to control, at least to have some understanding of what is happening, how it looks like, is it healthy or what, or not. So actually, let me show you one of the examples of how it looks like. So actually, we've built it ourselves. There are a lot of other companies and tools that provide the same kind of system, but I prefer to build it myself because I know what I'm interested for, what I'm looking for. So as you can see, we have the health of the system, we have number of sessions that were analyzed, the cost, we have the average scores and success rate, we have some trends, we have the AI insights, and this is the most important part when the AI and the agent try to connect the thoughts, find the patterns. If there are some critical ones, you have a description for each of them. What is this? Why it matters? What are the root cause? All the sessions affected? And some recommendation fix. So right now, this is just randomized data, but it comes from a real conversations. So we have a score distribution, each company, number of sessions or the cost you have. the sentiment analysis, some entities, we have analytics about tool calls, what was the success rate, how many rejected and why, and you have detailed analysis of each session. So it ranks each one of them, scores, it is how many times it was running, how many messages, how many tool calls, how many summary, and you have a detailed explanation for each of them, what was the problem. It is something that you need to understand and the limit helps a lot and it helps you to check and watch across hundreds of conversations. So the main goal of this dashboard and the system is not to fix some specific problems or bugs. This is more than a high level that helps you to watch across hundreds or thousands of real conversations. So you can run it once in a week, two times in a week or something like that. All right, so the next one is that, so the problem of the previous agent that I mentioned is like the first one helps you to detect the specific problems. The second one is to give you a high level understanding, but both of them work from an angle of the logs or a code or a session itself, but you need to have some kind of user perspective, right? So that's why we have some computer user agent that helps you to open the browser, to log in and try to simulate the customer itself because sometimes you have some problems in the UI. You need to check if everything looks good. And yeah, sometimes code and logs don't help you in this way. So this is an example how it looks like. So you use codecs sometimes. So it used the browser itself. Actually, this is pretty slow and we tried to build the specific skill that knows our website, our DOM, how it looks like and that is much faster and you can open the website, login, open the session, extend the messages, check what is happening and also check how it looks like, what are the artifacts and yeah, it was pretty good. But yeah, you need to... know that it will spend a lot of tokens. And also for all of those problems, you need to remember that you need to give access to all kinds of tools that is needed. So you need to have a log, projectors, a metrics, database, UI, so as the humans need, right? So you need to give all context, all possibilities to understand what is happening. For example, for the computer use agent, want to detect some problems. it should be able to analyze the trajectories, check the database to understand what happened. And, you know, that's why I'm calling it a mirror harness, the whole system when everything is connected and the PR or the answer from the agents are dependent on the, depend on the real problem and they're not guessing what is happening. So the mode is not the model alone. and you need to build the agent or a system or a harness around it which watch itself, understand, improve and help you at least to monitor what is happening, but also in the ideal case to close the loop, send automatic PR, notifications, wherever, and help you to speed up the process. So shipping is the easiest part today. If you want to build a production agent, you need to close the loop first because somehow people aren't talking about how important it is, what happens after you launch. So everyone can have the same model, everyone can have the same agent or harness, but you need to have some internal system that helps you in this process, detect the problems and give you a sense of what is happening and helps to make the product better. Thank you.

We run a production system of agents for real customers. The team that keeps it healthy is also made of agents.

Operating an agent product isn't like operating software. When our agent fails a customer — a dropped constraint, a stale price, a confident wrong answer — nothing crashes and no log lights up. The failure is in the conversation, not the stack trace. So we put agents on the operations:

- One monitors production conversations and judges where the agent actually let a customer down — across thousands of live sessions, not a sampled few.
- One watches logs and system health and traces real problems back into the code.
- One writes and runs tests, because "green CI" means nothing for a non-deterministic agent.
- One reviews every PR — human or agent-authored — against a single question: root cause, or just the symptom?

Humans stay at the merge and approval boundaries. The agents do the watching, judging, testing, and drafting that no human team could keep up with at this volume.

This talk is the honest version: what each operating agent actually checks, where we trust it and where we don't, what breaks, and why operating an agent system is becoming its own engineering discipline — done, increasingly, by agents.

Speakers:
- Raphael Kalandadze (Wandero AI): Co-founder and CTO of Wandero AI, an agent-native operating system for travel and hospitality, and co-founder of Tbilisi AI Lab, where we build the first Georgian large language model.
  X/Twitter: @RaphaelKalan
  LinkedIn: https://www.linkedin.com/in/rapael-kalandadze/
  GitHub: https://github.com/RRaphaellRaphaelKalan