One model, three roles — route by prompt

Background

The planner, executor, and synthesiser are the same LLM given different system prompts. The agent picks the mode by inspecting the system prompt:

• contains "planner" or "decompose" → plan
• contains "synthesiser" or "synthesize" → synthesize
• otherwise → react (the step-by-step executor)

Problem statement

Implement llm_mode(system_prompt) returning which role the prompt selects.

Input

• system_prompt — the system message string.

Output

Returns "plan", "synthesize", or "react".

Examples

Example 1

Input:  "You are a research planner. Decompose the question..."
Output: "plan"

Example 2

Input:  "You are a research assistant that works step by step."
Output: "react"

Constraints

• Case-insensitive substring checks.
• "planner"/"decompose" → "plan"; "synthesiser"/"synthesize" → "synthesize"; else "react".
• Check plan before synthesize before defaulting to react.

Notes

• "Multi-agent" here means different prompts and tools behind one interface — not different models.