The bigger the agent gets, the worse a giant prompt performs.

At some point, prompt engineering stops being enough.
What starts as one “smart” system prompt slowly turns into a brittle mess: expensive, hard to maintain, and easy to break.

That’s why agent design in 2026 is moving from prompts to architecture.

Instead of stuffing everything into one instruction block, it makes more sense to use small core prompts + modular skills + on-demand references.

The agent should carry an index, not the whole library.

https://medium.com/google-cloud/how-to-build-ai-agents-with-adk-and-skills-in-2026-from-prompts-to-architecture-2b7c5ff7003c