Similar thinking as mine (in the end i was long time enterprise architect responsible for f500 infra/app stacks:)))) and the swarm idea is something i will explore more in detail, it kind of fits in my thinking of the future of ai (call it agentic or not) where we will have multiple actors negotiating who will do the task at hand, for what cost and with its own requirements......and output precision/forecast of the outcomes:))))))
Please do share this around if you like it. I'm an independent writer with no institutional backing or support, so I rely entirely on word of mouth shares from my readers to grow my newsletter
Great article. @Devansh can you point to some demos ( videos are fine ) of any practical application of agents TODAY? +1 for digital marketing / growth use-cases. thanks and keep up the great work!!
Don;t know the Digital Marketing space super well, but I know a lot of RevOps people are looking into it. That might be worth exploring.
With something like Agents, there is a lot of misinformation/clickbait by people stuffing in keywords for simple things. My recommendation is to explore building your own. Let me know if you need help with that
Heh, was thinking that substack does not have anyone with real llm/ai build experience and here you are. Even though it might be quite heavy for most of the readers anyway:) great article:)
I'm glad you liked it. I think AI has that problem generally- most commentors don't actually build on the solutions so they don't have real insight into the tools/architectures.
If you like this, you might like these two articles as well-
Not intending to be critical of this article - which is both comprehensive and well-intended - what seems to be missing is a very important, crucial first step: design thinking. What problem are you solving and who are you solving it for?
You write: "Instead of overengineering a system, keep your first LLM Agent extremely minimal by augmenting your base LLM with the following:" - this is great advice. Yet, 99% of people are not even asking the question: what IS it for? Do I even need an agent to solve this problem?
I'm getting frustrated that agents are presented as the hammer and EVERYTHING, every single problem on earth, is suddenly perceived as a nail.
1) I'm not sure why you're commenting on about the importance of design thinking on a post "How to Build Agents". You wouldn't criticize an article on cooking fish by saying "Why is it not talking about making sure the reader can afford fish and is not allergic to them". If someone is on an article on How to Build Agents, the assumption is that they want to learn how to build Agents and it is somehow relevant to their work.
2) Agents are based on a simple design philosophy- don't use stochastic processes where auditable, deterministic (and generally more efficient) processes will work. If you don't understand why that's important I would ask you to show me a single modern AI system at scale (not just LLM based) that does not benefit from this philosophy. Agents extend that idea to LLMs specifically, since you get some interesting ways to work with them that don't exist in traditional ways (more diverse inputs).
3) "I'm getting frustrated that agents are presented as the hammer and EVERYTHING, every single problem on earth, is suddenly perceived as a nail."- show me one serious builder that has made this claim. Not Gary Marcus style claims in the air with no backup, but real quotes from actual builders (w/o perverse incentives) that say Agents should be used everywhere. The article I quoted from Anthropic, which has a LOT to gain by pushing Agents, literally acknowledges what you flag- "Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense."
If you're more interested in the specifics of Why Agents are so great, here is a detailed guide-
It wasn't so much a critique of your article per se - which I mention at the start. But I did miss this context of "Agents are based on a simple design philosophy- don't use stochastic processes where auditable, deterministic (and generally more efficient) processes will work" at the start of this post.
"Show me one serious builder that has made this claim." is a bit unfair, because you will argue that every serious builder 'gets' this, and the builder that doesn't get this isn't serious.
You can't deny that AI agents and Agentic AI is captured by marketing and sales teams across the globe to mean absolutely everything and nothing at the same time. I don't think a RAG bot is an AI agent and yet every idiot that can connect an OpenAI API + vector database calls themselves an AI Agent Agency. - Maybe that's what I'm ranting about and it is unfair to unleash that rant here (so apologies for that).
It's not as much the start of the post as much as it is a fact about Agents.
"because you will argue that every serious builder 'gets' this, and the builder that doesn't get this isn't..." -fair enough, but I think you'll have a hard time finding any builders who make this claim. The only people I've met that even think around these lines (and aren't profiting like Altman) are non technical people and they don't use terms like Agents. To them it's "AI".
"RAG bot is an AI agent"- I disagree since it matches this philosophy. It would be classified as a simple, workflow Agent.
Ultimately, everyone is branding themselves as such because there is a lot of low hanging fruit to be captured. Simple RAG and Data Processing can help so many important use cases. AI doesn't have to be complicated or "intelligent" to be useful to people. Hype is an unfortunate side effect, but that doesn't invalidate the philosophy or real developments happening here. Modern Agents are atleast 2.5-3 years old (became possible with GPT-3), it's just that people didn't give much attention to LLMs till recently.
I too prefer the planner/orchestrator subbing out to worker LLMs as required. Also… Zubimendi to City. Pep agrees with you.
It's one of the best ways to build things
Similar thinking as mine (in the end i was long time enterprise architect responsible for f500 infra/app stacks:)))) and the swarm idea is something i will explore more in detail, it kind of fits in my thinking of the future of ai (call it agentic or not) where we will have multiple actors negotiating who will do the task at hand, for what cost and with its own requirements......and output precision/forecast of the outcomes:))))))
the negotitation is an interesting idea
Stellar stuff here. Thank you for the time and energy you use to put this out to democratize the learning curve on this stuff!
Glad you liked it.
Please do share this around if you like it. I'm an independent writer with no institutional backing or support, so I rely entirely on word of mouth shares from my readers to grow my newsletter
@Devansh, We are building an AI Assistant/Tutor for K-8 in the US. Any comments/advice? Can you help or join? My email: costabrent@gmail.com
Sent an email
Great article. @Devansh can you point to some demos ( videos are fine ) of any practical application of agents TODAY? +1 for digital marketing / growth use-cases. thanks and keep up the great work!!
https://iqidis.ai/ is a fantastic example of using it for Legal.
Don;t know the Digital Marketing space super well, but I know a lot of RevOps people are looking into it. That might be worth exploring.
With something like Agents, there is a lot of misinformation/clickbait by people stuffing in keywords for simple things. My recommendation is to explore building your own. Let me know if you need help with that
Heh, was thinking that substack does not have anyone with real llm/ai build experience and here you are. Even though it might be quite heavy for most of the readers anyway:) great article:)
I'm glad you liked it. I think AI has that problem generally- most commentors don't actually build on the solutions so they don't have real insight into the tools/architectures.
If you like this, you might like these two articles as well-
Which Foundation Model is best for Agent Orchestration- https://artificialintelligencemadesimple.substack.com/p/which-foundation-model-is-best-for
6 AI Trends that will Define 2025- https://artificialintelligencemadesimple.substack.com/p/6-ai-trends-that-will-define-2025
Excited to hear what you think
Not intending to be critical of this article - which is both comprehensive and well-intended - what seems to be missing is a very important, crucial first step: design thinking. What problem are you solving and who are you solving it for?
You write: "Instead of overengineering a system, keep your first LLM Agent extremely minimal by augmenting your base LLM with the following:" - this is great advice. Yet, 99% of people are not even asking the question: what IS it for? Do I even need an agent to solve this problem?
I'm getting frustrated that agents are presented as the hammer and EVERYTHING, every single problem on earth, is suddenly perceived as a nail.
So many things wrong with this comment-
1) I'm not sure why you're commenting on about the importance of design thinking on a post "How to Build Agents". You wouldn't criticize an article on cooking fish by saying "Why is it not talking about making sure the reader can afford fish and is not allergic to them". If someone is on an article on How to Build Agents, the assumption is that they want to learn how to build Agents and it is somehow relevant to their work.
2) Agents are based on a simple design philosophy- don't use stochastic processes where auditable, deterministic (and generally more efficient) processes will work. If you don't understand why that's important I would ask you to show me a single modern AI system at scale (not just LLM based) that does not benefit from this philosophy. Agents extend that idea to LLMs specifically, since you get some interesting ways to work with them that don't exist in traditional ways (more diverse inputs).
3) "I'm getting frustrated that agents are presented as the hammer and EVERYTHING, every single problem on earth, is suddenly perceived as a nail."- show me one serious builder that has made this claim. Not Gary Marcus style claims in the air with no backup, but real quotes from actual builders (w/o perverse incentives) that say Agents should be used everywhere. The article I quoted from Anthropic, which has a LOT to gain by pushing Agents, literally acknowledges what you flag- "Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense."
If you're more interested in the specifics of Why Agents are so great, here is a detailed guide-
https://artificialintelligencemadesimple.substack.com/p/why-is-agentic-ai-so-powerful-agentsthoughts
If it still doesn't convince you, I don't know what to tell you.
It wasn't so much a critique of your article per se - which I mention at the start. But I did miss this context of "Agents are based on a simple design philosophy- don't use stochastic processes where auditable, deterministic (and generally more efficient) processes will work" at the start of this post.
"Show me one serious builder that has made this claim." is a bit unfair, because you will argue that every serious builder 'gets' this, and the builder that doesn't get this isn't serious.
You can't deny that AI agents and Agentic AI is captured by marketing and sales teams across the globe to mean absolutely everything and nothing at the same time. I don't think a RAG bot is an AI agent and yet every idiot that can connect an OpenAI API + vector database calls themselves an AI Agent Agency. - Maybe that's what I'm ranting about and it is unfair to unleash that rant here (so apologies for that).
It's not as much the start of the post as much as it is a fact about Agents.
"because you will argue that every serious builder 'gets' this, and the builder that doesn't get this isn't..." -fair enough, but I think you'll have a hard time finding any builders who make this claim. The only people I've met that even think around these lines (and aren't profiting like Altman) are non technical people and they don't use terms like Agents. To them it's "AI".
"RAG bot is an AI agent"- I disagree since it matches this philosophy. It would be classified as a simple, workflow Agent.
Ultimately, everyone is branding themselves as such because there is a lot of low hanging fruit to be captured. Simple RAG and Data Processing can help so many important use cases. AI doesn't have to be complicated or "intelligent" to be useful to people. Hype is an unfortunate side effect, but that doesn't invalidate the philosophy or real developments happening here. Modern Agents are atleast 2.5-3 years old (became possible with GPT-3), it's just that people didn't give much attention to LLMs till recently.