# Separating Agent decisions from LLM generation in system design

A guide to separating agent decision-making from LLM generation, using fast decision layers for routing, scoring, and gating in a testable architecture

Grade: X — Not a typed-decision job: needs written output, plain rules or lookups would do it, or needs raw image or audio. Also marks ideas that must not be built.. Score: 3.61.

Vertical: Software & tech.

Function: Agents & dev.

Status: Seen in the wild.

Volume: routine.

Value: meaningful.

Risk: low.

Evidence: described plan.

Flags: check-fit.

- [Source](https://madewithjev.com/community/jev-agent-engineering-separate-decisions-from-llm-generation)
- [Vertical](/verticals/saas_tech/)
- [Function](/functions/agents_dev/)
- [Prompt injection detection guardrail](/use-cases/prompt-injection-detection-guardrail/)
- [LLM output safety classification](/use-cases/llm-output-safety-classification/)
- [LLM cost-quality routing with RouteLLM](/use-cases/llm-cost-quality-routing-with-routellm/)
- [Claude Code hooks pre-tool-use validation](/use-cases/claude-code-hooks-pre-tool-use-validation/)
- [Log sensitive-data leak](/use-cases/log-sensitive-data-leak/)
- [Prompt injection scanner](/use-cases/prompt-injection-scanner/)