Designing AI agents is easy... until you want to do it right

I know, there are 345 millions videos (educated guess :P) on how to create your agent with framework Z, Y and X.

Just a couple of lines, e.g. ai-sdk.dev:

import { ToolLoopAgent } from 'ai';

const myAgent = new ToolLoopAgent({
  model: "anthropic/claude-sonnet-4.5",
  instructions: 'You are a helpful assistant.',
  tools: {
    // Your tools here
  },
});

That simple, or as Strands docs depictures this:

Pretty easy, huh?

Yes, until you want to do something series, or has a fetish of doing things in a right way, not just to show that you have "AI" in your CV.

When you try to do so, where are a couple of uncomfortable questions that arise, like:

• how we should connect to LLM?
• and do we want to have only one or multiple?
• how we should structure the prompt to execute in a way we expect it to execute?
• ...especially with different LLMs?
• how to balance tool calling with internal model reasoning? (
• what if it doesn't like the way we expect it to call our tools? (or maybe doesn't support tool calling at all as some Ollama models)
• what if where are too many tools and model gets confused and user gets infuriated about token usage?
• what happens if the user sends another message when the current one is still processed?
• what if the user cancels the execution?
• what it there is a tool executing in that exact moment?
• how we want to bill the user for that execution of tool and LLM when it was cancelled?
• let's do a step back - how we want to do a billing and cost control for the users in the first place?
• even if you figure it out, what happens when you have more than 10 users? I'm not afraid about your Kubernetes cluster or Lambda, we both know that they can handle not 10 but even 20 users of your application, how about your provider for LLM? (and this is not something I got out of the blue, just google for... Gemini tiers :D)
• then... let's open a Pandora box of MCP servers, which is similar cancer that JavaScript was (is?) in 2000's and 2010's

I can go on and on with hundreds problems like that - I'm kind of old-fashioned, I did "Kubernetes the hard way" so did I with agentic AI systems during last two years.

If you look for easy ways and quick business value - just close this page - sorry, it won't be a nice journey, but I hope the insightful one :)