The Agentic AI Engineer - Benedikt Sanftl, Mutagent

Okay.

Hi, everybody. Uh, welcome to our talk: the Agentic AI Engineer. I'm Bene, CEO and Co-Founder of Mutagent, and I'm here with my colleague.

Hi, I'm Burak. I'm the CTO of Mutagent, and today we're basically going to talk about loops and how the Agentic AI Engineer works.

So, as you're all aware of now, loops is the hot topic: how you build software in an agentic loop, and we apply the same loop to the building of AI Agents. And, as you're all aware, there's, uh, two concepts here.

One is the offline loop, where while you build, you iterate on your agent: you test it, you evaluate it, you improve it, and you go on. And then you have a second loop, which is what we call the online loop, where once your agent is deployed to production, you monitor its traces, you diagnose it, and then you feed it back into your optimization loop, yeah, to iterate and have multiple versions of your agents.

Yeah. Up to until now, what we did was this loop, doing this loop manually. It's quite slow. The life cycle is basically: you have an issue, you want to change something to your agent, yeah, you implement the change.

You maybe vibe-implement the change if you use coding agents for it. Yeah, you generate some samples for this new feature or issue to test it. Yeah, then you look at the result, you look through the traces, how does the outcome look like.

Then you maybe ship it, you do A/B testing, and all your feedback is kind of manually, it takes very long. Yeah, and the bottleneck basically becomes the human review and the human, yeah, building time. And, yeah, that you can't scale, especially if in your organization you're now planning to roll out hundreds of agents, etc.

Yeah, and, yeah, this is why we think the Agentic AI Engineer is the natural next step to build agents. And I'll have Burak deep dive into how we improve timing and the road to production reliability with the Agentic AI Engineer.

Uh, so, yeah, the key thing here is basically once you reach a certain number of agents or AI-based features, the human performing this loop, again, cannot really scale in enough time. So this is why doing this agentically is the key to increasing the throughput, because then you can fit many more cycles into the same time window.

And now, how that loop works is basically we have a few stages. So this is when you are starting from scratch, like the current software development practices, you first create a spec for your agent, or your skill in this case, and here you need to define all the responsibilities and the functions that agents need to handle, the decisions that it has to make on certain conditions.

And here, again, this is only the definition stage. Once you defined your agent's requirements, then you can finally go on to the build. And build is where you then realize that spec in a specific harness or agent framework, or in these days you could even build it as a Claude code or a Codex agent.

Then comes the next step: this is where you define clear evaluations to evaluate your agent's performance, because these are the key metrics then where you can say, hey, my agent is functional or not. You can think of essentially equivalent to unit tests for coding.

This is how you verify your agent works. Then, after evaluation, if everything looks fine, you usually have the ship, basically where you deploy this agent to production. Again, this can be a code update, this can be a direct update on any agent platform, or again, your local harness agents.

Then comes the online part. This is where then the agent is continuously monitored for issues, and based on certain trigger conditions, then you can start automatic diagnostics. Again, this can be based on the volume of traces that your agent generates, or, you know, weekly or daily jobs.

Then we go into diagnosis stage. This is where you collect all the failures for your agents and do structured root cause analysis to then understand where the failures are coming from. Once you understand and categorize the failures, then you can finally go on to the optimization stage.

This is where then you create, let's say, very specific changes or mutations for your agents to deal with the found failure modes. And then the whole cycle repeats again, you evaluate, and if everything looks good, then you can deploy again.

Now, we will maybe do a deep dive on each stage, what that entails.

So, before we continue, Burak, we have two paths here. Like, one is the cold start path, and one is based on existing features. Should we dive deeper for, like, a half a minute on what difference is here, what we see, and why this is important?

Yeah.

Yeah, so today, if you again sit down to build an agent, there are, you know, two options. Either you already have an agent, it's built, it's already running somewhere with, you know, certain accuracy. However,

the other option is, again, you are creating from scratch. So when you create from scratch, obviously you design with the spec and the conceptualization stage. If you have an existing feature, the most likely, again, the agent is there, but then you are optimizing over something that already exists.

Okay, cool. And, yeah, let's dive deeper into each phase. I mean, you just mentioned the speccing of new building agents, so this seems to be an important artifact, so let's dive into it.

Right. So with spec-driven development, also prevalent for building software artifacts these days, or any kind of coding agent workflows, basically the goal is to capture the requirements for the agent, and especially the success criteria. The reason being, apart from coding agents, the agents in other domains, they handle specific processes, and this differs from company to company, so it's very specialized to the environment where the agent is in.

And then the key point here is to clearly define which context requirements that the agent has, again, which integrations and tools it then needs to have. What are the jobs to be done, or the responsibilities that the agent will handle, and what it will not.

And then finally, again, the constraints, and in general, the boundaries for the agent.

Okay.

Now, once we can define a clear spec, as I said, it becomes like a blueprint for any future development, which then the implementation is held against.

Okay. Then let's dive into how we would then build an agent from the spec,right?

Right. So basically then the spec tells your coding agent

what to build, and then here the target platform choice is entirely yours, because as I mentioned, the agent space is very changing very rapidly these days. So the framework you are using today, you might want to change, you know, in a year or so.

So essentially, because of that, the spec is

isolated from the implementation details. So once you decide to build, you can pick any target. Here then your coding agent of your choice will take that spec and give you a, you know, initial version of that agent, which is then basically customized to run on any platform that you see fit.

Why would I want to change the platform in a year down the line? So what did we learn out of our experience here?

Again, building agents for the last 3 years, sometimes the agent framework or the harness does not always have the capabilities, so occasionally you hit like a bottleneck or a roadblock, and then you have to rely on the underlying framework to kind of get rid of that, and this can sometimes take a while.

The key here is to be flexible, because, yeah, essentially, again, you want to pick the best harness or the framework that can, you know, fulfill your requirements.

I mean, we've all seen it in the last months, how new harnesses shipped and, yeah, how everything went from, like, agents building code towards, yeah, defined as like an agent loop runtime. I mean, we've seen Hermes coming up, deep agents, and all of these frameworks,right?

Yeah.

Okay, let's continue. What happens after build?

Now, after you build for agents, you essentially go into the Eval-Driven Development Loop, which I would call, and this is kind of equivalent to test-driven development for building software with agents, because then the agent needs a termination condition,right?

So when is an AI feature or an agent good enough? So here, again, there are two ways to kind of create your eval suite, which is composed of the evals, like the metrics and the criteria you evaluate, and the data sets themselves, which usually contain the cases you have to satisfy.

Now, in the beginning, the original option is that you can sit with your domain experts and then try to write down eval metrics and criteria that would cover the feature or the agent you want to build, but most teams working on that will already know that this is a bit difficult, as in you cannot pre-guess the entire evaluations suite from the beginning.

And secondly, you can always start with historical data or synthesized data from a known, you know, sample of the data that you would like the agent to be tested on. But essentially, the real and the complete eval suite is a product of discovery.

What that means is, over time, from user feedback, from production failures, you collect the metrics and criteria, plus the additional data set cases, which is often representative for edge cases or, you know, hard cases that the agent needs to deal with.

And with that, then you finally have an evaluation suite where you can run the agent against and exactly know where it fails.

Okay, before we continue, why does the Agentic AI Engineer help us here so much? So obviously we're talking about the evaluator agent, so why is it so good to use this kind of concept or thought process in the building?

I mean, one issue is, again, imagine you have a data set item of 200 here without automated evals running this. And evaluating by, like, human eyes takes quite a while.

You would have to, again, scroll through an observability dashboard and logs, and this in turn increases your loop time per eval state. So as soon as you have a lot of features that you need to evaluate and experiment on, then it suddenly becomes impossible to, you know, do this quickly or in parallel.

Then the human essentially becomes the bottleneck.

Okay, so we just have an agent do this work and sifting through traces?

Yes, as the current era says a bit, you know, you design loops for your agents so then they can autonomously work as many of these things in the background, and then your job becomes designing these loops with a clear eval or termination gate.

Okay, understood. So let's look at how a good eval is supposed to be constructed and what it kind of evaluates.

Right. So in the context of agents in general, mainly the trajectory is important because agents receive input or, let's say, intent or tasks, and then they have a specific system prompt or like a decision tree that they kind of operate over.

And when doing agent evaluations, again, in general we check, hey, was the context complete, as in did the agent have all the required context to perform the task end to end? Then this includes also checking every, you know, tool output in the trajectory, because every wrong tool output in a session can in the end lead to a wrong output as the final output.

And apart from that, again, there are different things you can evaluate on. So these days, also the harness that the agent is operating on has quite drastic effects on the agent behavior, and this is, again, another vector of optimization.

But in the end, when you evaluate an agent, you would like to evaluate all of these things and not something in just isolation.

Mm-hmm.

Right. And in general here, what makes an eval useful? So you can always have, you know, metrics or evals that are working in an LLM as a judge fashion and give you some score. But here, in order to make an eval, you know, useful, it has to provide actionable feedback, as in

when you use score-based evals, unless your rubric is very well defined, then this does not exactly tell you what to fix. In such cases, using binary type of evals or criteria is preferred, because there you have a kind of a call to action.

If an eval or a criteria fails, then you know exactly what happened and how to kind of deal with that problem. Then another point is your LLM as a judge solution should be calibrated so that you don't have the scoring noise between judges, because since LLMs are undeterministic, what you will mostly encounter is the same judge can evaluate a problem different ways on each run.

And then here you have to make sure your LLM as a judge solution deals with this variance problem, otherwise it's hard to run experiments and conclusively say, hey, my initial or my improved version is better than my initial version of my agent.

Mm-hmm.

Right. After then evaluation, basically we go live, and this is where then collecting, let's say, learned failures, failure modes, or primary signals from the executions takes place. Here the idea is, again, when an agent encounters a problem, over time there will most likely be multiple occurrences of these.

So it starts by identifying what failure mode that the agent has encountered, and then grouping essentially these failure modes by their root causes and where they originate from. Again, this could be a section in the agent prompt, this could be missing tools or, let's say, malfunctioning tools.

But after your diagnostics results are categorized, then you can finally generate new evaluations, first to detect these problems, and then second, you can generate improvements and remedies based on these problems.

And yeah, with that, over time you have a buildup of these learned failure modes. So then every agent over time gathers these historical data that it can always check against. When you start diagnosing, there's a bit of an upfront cost in the beginning, because you often need to deep read the LLM traces to find out what's going on.

Over time you can collect code-checkable indicators per failure mode. What that means is there are maybe specific pieces of content or there's a specific tool call sequence where you know that the agent will encounter an issue, and then this later on helps you to diagnose problems in your traces without actually having to read through all your traces.

Now, why is that a problem? Because if you have, let's say, millions of agent traces,

trying to read all of these actually costs more than the execution itself, so it's not the most efficient way of diagnosing the problems. What you want to aim for is, again, try to pick a representative sample from your whole traces by, you know, intelligent segmentation strategies and using, again, as I said, the learned indicators.

Okay, cool.

And then with that, you can finally start building the autonomous optimization loop, because here then, given that you have an eval suite, you can, you know,

vary your feature, again, update certain sections or even run auto-research style experiments to see that whether you can reach the desired or the target scores for your evals. As long as you can reach that, then it's automatically shipped to production.

Then from production you have your outer loop. Again, every issue that's found then on diagnostics gives you an improvement or a remedy which then you can optimize on, and then as long as the eval suite is green, then this, again, gets deployed to production.

Let's continue now with the whole life cycle. Yeah, as we learned now, the details of each, everything starts with a spec. You define and design it, you build your agent, and all of this can be done obviously agentic.

So you define agents that do this work, and then they all together become the Agentic AI Engineer. You go through the offline loop where once you build your evaluation system, yeah, you evaluate, you optimize, you test, you improve your accuracy on your test data set.

Once you feel ready, you deploy it to production, and that's when the online loop starts. Here you get real feedback from users. Yeah, real test cases. You continuously monitor them, you look at your live traces, you diagnose them.

As we learned before, with a diagnosed agent you can do root cause analysis, you can cluster your failure modes, you'll derive new eval criteria from it. They become part of the spec, they become part of the agent, so they continuously grow with you while you use your agent in production.

And yeah, this becomes the online loop, and the more use cases you collect, the more production data you see, the better your agent becomes and the better scoring your agent becomes. And this all together now becomes one end-to-end loop that you can run agentic.

And yeah, from here on, yeah, you can now see the combination to software-driven development with coding agents. You can transport these concepts also for AI Engineering and building AI agents. And now, as we at Mutagent work on this, we're going to show you now how this looks like as a product.

This is kind of how it looks like. For now, everything runs in your environment, and as cloud and local, we'll offer and manage servers down the line. So it's a set of agents. We have two in research preview, the ones we talk deeper about this in our talk.

First one is the evaluator agent, it helps you build an eval set, a good data set, because this is the core of the optimization, the eval-driven development loop. The second one we have is the diagnose, the diagnostics agent.

It does you, it helps you diagnose your traces you already have in productions, because we learned from our users that reading through these traces took them a lot of time and having kind of an agent they can just spin off from their coding environment to analyze their traces is very helpful to them.

And as you can see here, both of these agents are connected through an orchestrator which runs in your coding environment, and what you need basically is connectors to sources where you basically have all your traces, where you get your incidents from.

This could also be like a ticketing system, Slack, where people report failures of your agents. Then you connect them, and then obviously, as we mentioned before, there's different target platforms. This could become a PR that automatically gets raised in GitHub.

This could be just the adjustment of your agents in, yeah, in MD files. Yeah, different frameworks we target, and/or being deployed to manage services. Yeah, kind of this is how our platform works. We spin up different agents, they're all connected to an orchestra.

And I guess, Burak, let's show our first agent in research preview to the audience.

I. It in practice, as I said, you have an orchestrator which kind of handles the dispatch for all the other sub-agents or the stages we have. Then you can use it in any coding harness or agent you want.

In this case, I'm using Claude Code, but then what it will do is when you first boot up, it will show you kind of like a dashboard of which stages you have and also kind of the existing things in, let's say, your code base, your agents, and the prior configurations.

Now here I have some star commands. These are basically, let's say, trigger commands for the workflows or specific stages. And if I type diagnose here, this will basically start the diagnostic stage for the root cause analysis. And here, again, you can point the diagnostics into a specific scope.

This can be an agent, this can be a skill as well. So you could basically diagnose all invocations of a particular skill. Here,

when you start the diagnostics, it will retrieve traces from your configured source platform. In this case, again, it can be something like NlangViews, it can be your local Claude transcripts or another, you know, like a JSONL format that's like exported from one of the observability platforms.

Now, the diagnostics run itself takes quite a while to finish, so I will show you instead a pre-generated kind of version of what it does and how that looks like in the end.

Okay, so what do we see here now?

So basically, once you run the diagnostics agent, as I said on your agents, you will get a generated HTML artifact which basically should show you, you know, the details of your agent, which tools you have. Then in general, if you have code access, it will also tell you which harness or the framework then the agent is using.

And then here in general, it will basically go through your traces and then select primary signals or, let's say, failure modes depending on how frequent they occur in the trace samples. Now, as I said, most of the time you don't want to read all of your traces because this is not cost efficient.

So there the diagnostics use like a multi-tier filtering or segmentation to kind of pick a representative sample. So originally a few, you know, pieces of the traces or, let's say, a portion will be read by the LLM to see that if there are any detectable obvious problems.

And based on that, then the diagnostics then decides or to focus on a particular failure mode or a signal. The other option is then when you trigger the diagnostics, you can also specify an issue that you have seen yourself or issue that you're looking for.

This could be something that the users reported. In this case, it's again more of a guided search, meaning, hey, if you tell the diagnostics agent you have a particular problem you're interested in, then it will try to find all incidents or occurrences regarding that same problem.

Okay, and what do we see here in, yeah, in full here?

So this is in general the overview which tells you again which issues kind of were detected and kind of shows you like the frequency in the given timeframe. Then for each tab, you will have a failure mode and kind of like an explanation of the problem.

And then here you can kind of see where that problem comes from or why it happens. The agent will provide kind of like a recursive why chain, let's say, to then tell you, hey, this is where the issue is originating from.

Now, in addition, it will offer you kind of remedies or fixes to deal with that. And then here you will see an assumptions block. Here, this is important because when you are reading traces, sometimes you don't always have access to the code.

So this helps us detect

when the LLM or, let's say, the diagnostics agent makes certain assumptions that are also not correct, and we can see and correct them here. But in the end, you will be presented by certain corrections or remedies which then kind of can fix your problem.

You can, you know, pick either the recommended or, you know, as many as you need as it's multi-choice. And once you basically go through all your problems, at the end you kind of get to a decisions page. And here again, for those using, you know, text-to-speech or speech-to-text AI features like Whisperflow, there's always a general feedback box so that you can always talk into your mic.

And finally, when you are happy with all the decisions, you get a markdown, let's say, task definition for your coding agent. So all the fixes or the remedies can then be directly applied once you go back to your terminal and to your coding agent.

Thanks for the demo, Burak. So

last words on our product?

Yeah, so thanks for listening into our talk. Yeah, I hope we could show you some good insights on how we envision the future of agent building with the Agentic AI Engineer. Yeah, stop the debugging, have agents do the tedious work, and reach out to us if you have further questions.

And yeah, have a great conference.