lilfrogdev

I've been thinking a lot about what it means to be LLM agnostic.

Not in the abstract "all models are interchangeable" way, because they aren't. Each model has its own strengths, weird behaviors, pricing, context limits, and other weird quirks.

But I don't want my workflow to depend too heavily on a single provider.

Why

It's easy to get attached to whichever model feels best this week.

One week it's Claude. Another week it's Kimi K2. Then a new Codex model drops, or Gemini gets better, or a super powered "Mythos" appears and isn't accessible to the public.

If my workflow is built around one provider, switching feels expensive. I have to install a new cli, relearn commands, and accept whatever limitations that provider has.

I'd rather build a workflow where the model is a piece I can swap in and out as needed.

Different models are good at different things

I think of LLMs a lot like expensive programming languages.

There isn't one language that's best at everything. Python is great for quick scripts and automation. Rust is great when you care about safety and performance. TypeScript is great for building web apps. Bash is ugly until it's exactly the right tool.

Models are the same.

That means the question shouldn't always be:

What is the best model?

But instead you should ask:

What is the best model for this part of the work?

You wouldn't want to use Opus 4.7 just to change a variable in your project. Instead you should consider a free open source model or a local LLM.

Avoid provider loyalty

I don't think provider loyalty helps much.

Model quality changes too quickly. Pricing changes. Session limits change. Product direction changes from consumer to enterprise. Sometimes the best model for your workflow is just the one that currently fits the task with the least friction.

Being loyal to a workflow feels more useful than being loyal to a provider.

For me, that means:

keeping my editor and terminal as the main pillar
using tools that support multiple providers (ex: Pi & OpenCode)
comparing models by task, not hype
be willing to learn and adapt to the rapidly changing environment

The goal

The goal isn't to use every model.

The goal is to avoid being trapped by one.

If a model is great, use it. If another one is better for a specific task, use that. If a provider changes direction or gets worse, your workflow shouldn't fall apart.

LLMs are moving fast. The best setup is probably the one that lets you move with them.