Hoje em dia escuto A Hora que é o antigo Foro de Teresina e o Medo e Delírio

você é o pig em pig na cidade grande e não sabia

Esse ficou meio apagado comparado aos outros

The idea behind “agentic” workflows is that your main agent dispatch other agents (sometimes with smaller models) with smaller context so they are cheaper and faster. Models can also degrade on larger contexts (they forget earlier context tokens)

The context window is a limitation on the model/capacity, nowadays most frontier models have a 1M+ token context window compared to 100/200k tokens from earlier. They compact the conversation when reaching ~80%, but some context is lost and the model can feel “dumber”.

It is, but it’s standard to count cached vs non-cached tokens on different price tiers.

Lembrei do Decotelli

sem calcinha*

esperando