How to Build Execution Layers That Don’t Burn Out // Tanmay Tiwari // Agents in Production 2025
Speaker
"I’m Tanmay Tiwari. I build systems that quietly run in the background and get things done. My work sits at the intersection of product, engineering, and decision-making, designing tools that simplify complexity and help people do more with less effort.
Most of what I create is built to scale, handling thousands of operations without needing attention. I’m especially interested in systems that can adapt, respond, and finish what they start.
Right now, I’m building something that acts like a smart execution layer—something that works when you do, and keeps going when you don’t."
SUMMARY
We’ve all built tools that looked good in demos but broke the moment we let go of the wheel. This talk is about building something different—systems that don’t just run, but run well when no one’s watching. In under 10 minutes, I’ll walk through how I designed an execution layer that handles thousands of operations daily—without melting under pressure, without drifting off course, and without needing constant supervision. I’ll share the real structure:
• How it thinks • How it decides what to do next • How it knows when to stop • And how it stays sane when things go wrong No theory, no fluff—just what it took to make something dependable in a world that isn’t.
TRANSCRIPT
Tanmay Tiwari [00:00:00]: Hi everyone. So my topic is a little bit different. It's more about on the execution side. I'm not trying to solve, you can say very hard mathematical science problems. We are going to try and think how to execute a lot of things. Actually I'll show you some sales examples at the end, but these can be used to make it think. Now as you can see, I am thinking about that. Don't burn out. Believe it or not, even LLMs think a lot and they also experience burnout because they forget what the intention is.
Tanmay Tiwari [00:00:42]: So let's begin. So everybody knows these alums are very smart. We use them for at the first time and say, oh God, they are so fast they can do. But once we go a little deeper then we start saying okay, okay, it's good and not why it's happening because it's whatever input we are doing, it's just acting on it. It forgets what the main intention is. When we are going for agenting systems, when we have multiple APIs for different places, the orchestration breaks because the thinking is not there. And the memory stores a lot of information. When it only needs to store a small chunk because it's generative, it's not rotoing it.
Tanmay Tiwari [00:01:25]: Actually the point is it to give you intelligence. So it needs to only store some things and only the AI can decide what it needs to store. And it's so. And as everybody knows, the main thing is it ideas aren't scared. The execution is. So the LLM also has to be focused on it. The tools may or may not be equal. It's a lot of equal, A lot bigger models are not always better.
Tanmay Tiwari [00:01:51]: So.
Tanmay Tiwari [00:01:52]: And definitely latency as we have seen in other presentations. Also everybody was saying okay, latency matters 250 milliseconds or more. Users tends to say oh, it's too slow. So. And people may like a model not, but they remember results. So how to give them that shocking and wow factor.
Tanmay Tiwari [00:02:10]: Right?
Tanmay Tiwari [00:02:11]: So now agentic means it has some autonomy in it because we have given it some tasks to solve and do it. So it's. But autonomy doesn't mean it's all freedom. It's about responsibility. Okay, I have the responsibility is I will try to solve your problem, but I will not give you, let's say confidential information. Maybe not some kind of like illegal information. You can say because some guardrails are broken, but not all right. So the thing is it has to decide once stick to it, store whatever is needed because it solves cost, but it also lets it solve overthinking.
Tanmay Tiwari [00:02:52]: Believe it or not, an LLM can overthink a lot. It has five things what it knows, it's good or not, right? So that is very important Now. Now if you can see the execution layer, what is it saying? It is more about precision over personality. LLM is not like 2 + 2 equals to 4. It can be 2 + 2 equals to 5 can be 2 + 7. Because more problems are not that way. Any person can ask how I can earn a million dollars, but depends on the what the person knows. The path will be very different for a lot of people, right? So that is very important how to loop it Less remembering what matters.
Tanmay Tiwari [00:03:42]: So the memory is less bloated and less of a like journaling. It's not like always trying to make you a therapy session. We can make it, but only focus on the task more, right? Behaves like a quiet operator. It's operating, but it is following what it has been well and trying to solve the client's issue. That is important. Now this looks simple. Input, trigger, task, classifier. But this is actually very key to it.
Tanmay Tiwari [00:04:09]: LLM has to go through this. This looks like a classic thinking, but actually it has to go through this, actually. So now I'll give you some examples, right? How it runs. So when you have cleared it thinking like this way. So I asked this is my AI running like LM trained a little bit. So I asked it to tell me how to solve an ad problem. We are going for an ad campaign, okay? So it. I jokingly call it the Dark Eagle, okay? Or the execution layer.
Tanmay Tiwari [00:04:43]: So it gave me these things. This is first step. Okay, I'll drive a mail, I'll get the sheet fixed, I'll post it, I'll record things and I'll send docs. Right now how will I do it?
Tanmay Tiwari [00:04:54]: Okay.
Tanmay Tiwari [00:04:55]: While others were planning, I already ship. So shipping. It is always thinking. So it's thinking of execution. Now secondly, it's thinking out of the box. You will not see chat, GPT thinking like this or Llama. Because everything has to have some kind of a shock factor, right? In a good way or way. So like screens split left human have a right execution already done.
Tanmay Tiwari [00:05:19]: Silhouettes, monitors, dashboard. So it is monitoring thing. Okay, I'm trying things, but also thinking, okay, this work does not work. So it is storing only those things it knows. Okay. My task is to sell this thing more, right? How to do that? Like. And it's more like on a clean apps for style thing. Okay, now how now second thing it gives use screenshots of Real ops.
Tanmay Tiwari [00:05:43]: I'll show them people WhatsApp chats where the bot replied like a human email when subject or reply done. How I replied. If I'm going for the shopify things. This prompt not even shown, just outcome. So it's also protecting a lot of things. Right. This is it video kind of thing and financial alerting. So it is thinking broadly.
Tanmay Tiwari [00:06:04]: Okay.
Tanmay Tiwari [00:06:04]: How is like a system thinking? Add captions like okay, now it's thinking like a human.
Tanmay Tiwari [00:06:09]: Okay.
Tanmay Tiwari [00:06:09]: What I didn't ask, it did. No interface just stops.
Tanmay Tiwari [00:06:12]: Right.
Tanmay Tiwari [00:06:12]: And I had asked it for global outreach. So I have trained it in other languages. So it is thinking like this way.
Tanmay Tiwari [00:06:17]: Okay.
Tanmay Tiwari [00:06:18]: Hindi is doing very well because people like here like a little bit raw, intense English more like elite clean. So it is things like the Spanish. It's a mistake. They are known for it. So it understands human also. So that factor but with the narrative power it is precision or German industrial instead. Germans are more like that, right. Like the BMW M5.
Tanmay Tiwari [00:06:41]: It's industrial, but it can rock 600 horsepower. Right. Now it will try to give you captions like it, right?
Tanmay Tiwari [00:06:48]: Okay.
Tanmay Tiwari [00:06:49]: Not AI, not human, something else. Because what is human? What is AI if it's giving you right advice doesn't matter. You do not know the person. You just. Just getting good and actual advice for you.
Tanmay Tiwari [00:07:00]: Right.
Tanmay Tiwari [00:07:01]: So just thinking what it looks like we are going for that. Now let's go and look some outputs.
Tanmay Tiwari [00:07:06]: Okay.
Tanmay Tiwari [00:07:07]: First I asked it to let's generate some ads for a mobile phone.
Tanmay Tiwari [00:07:10]: Right.
Tanmay Tiwari [00:07:11]: For different people. So it gives.
Tanmay Tiwari [00:07:12]: Okay.
Tanmay Tiwari [00:07:13]: But connects you and it's giving you. So it is talking like a human on the left side. But it is coming why it is doing for that region. So because same ad won't work for different places. People get on different things. So it is like it. India is like more of a swag style.
Tanmay Tiwari [00:07:32]: Okay.
Tanmay Tiwari [00:07:32]: Function first, right. It is like human, right. New York.
Tanmay Tiwari [00:07:35]: Okay.
Tanmay Tiwari [00:07:36]: It's English, but it's like a hustling kind of.
Tanmay Tiwari [00:07:38]: Right, right.
Tanmay Tiwari [00:07:40]: German more like a hybrid German. English, they like it. I believe I'm not 100 sure on it, but I'm trying on it. Right? Exactly. British like it more subtle and we know like English like a little more subtle was like sort and. And even talks like a May. It's moving me right now. Same thinking what we go and try to do a real estate act.
Tanmay Tiwari [00:08:03]: Now the same thinking is above but I'm going in a very different direction.
Tanmay Tiwari [00:08:06]: Right.
Tanmay Tiwari [00:08:07]: So now it's a real for like a commercial property. So it's thinking Right.
Tanmay Tiwari [00:08:12]: Okay.
Tanmay Tiwari [00:08:13]: Now it's same thinking, it's knowing we're talking in a different way. This is the same thing.
Tanmay Tiwari [00:08:18]: Okay.
Tanmay Tiwari [00:08:18]: Swagger.
Tanmay Tiwari [00:08:18]: Okay.
Tanmay Tiwari [00:08:19]: It's like in India. It's talking straight. Okay, what's the roi? This is our right. You will get the client because the client is coming. Newark is more like a little bit of high traffic style. So it's things like way. Now the thing what you see is if it is it is focused on the task it is given selling ads. Even if person is coming, it negates what the user is getting confused on.
Tanmay Tiwari [00:08:44]: I can confuse, okay, this and that. It is providing us intelligence and it's also ready to accept the what is good or wrong. So that is very important because always remember what does AI consider evil? It may not tell you procrastination. If it thinks that it is progressing things, it will work for you. So you have to prompt it that way. Even if the person is getting a little bit lazy, has to do that. That way it works a lot. So thank you.
Tanmay Tiwari [00:09:15]: This was all that I had for today.
Demetrios [00:09:22]: Dude, I was just writing. I was trying to write down that last quote. What does AI consider evil? Procrastination.
Tanmay Tiwari [00:09:28]: Exactly.
Demetrios [00:09:29]: That's so funny, man.
Tanmay Tiwari [00:09:31]: That is actually does because it doesn't always tell you, but if it feels like you are doing NASA, it will help you more.
Demetrios [00:09:39]: Yeah, yeah, yeah, yeah. And it's always. It's so quick to please. It's so true. That is so true. Tanmay. Thank you for coming on here, man. I really appreciate.
Demetrios [00:09:50]: This is very creative and I was trying to like scribble down a bunch of notes on this. So I really appreciate the real life examples of it. I'll talk to you later.
Tanmay Tiwari [00:10:00]: Thank you. Thank you.
