This section covers everything that tells an agent what to do — its always-on identity, the distinct journeys it can run, and the reusable prompt snippets it can call on for specific jobs.
Overview
Instructions is the agent's core, always-on guidance: a role, a goal, a detailed instructions body, and the conversation starters users see first. It's compiled into the system prompt on every reasoning step. Tasks layer on top of that — distinct journeys the agent can run, each with its own objective, step-by-step instructions, allowed tools, and optionally its own model, so one agent can handle several different kinds of work predictably instead of relying on one giant instruction set. Prompts are separate again: named, swappable templates for specific functions like response generation or rephrasing, reusable and testable on their own.
What's in This Section
Topic | Covers |
|---|---|
Instructions | Role, goal, instructions body, and conversation starters — how they become the model's system prompt. |
Tasks | Defining distinct journeys, tool selection strategy (Routed vs. Direct), and task state at runtime. |
Agent Prompts | Creating, testing, and managing an agent's reusable prompt templates. |
How These Three Differ
Think of it as a hierarchy of specificity: Instructions shape everything the agent does, all the time. Tasks scope a subset of that behavior to a specific journey, with its own tools and rules, active only when that journey is selected. Prompts are the most narrow — a single named template swapped in for one function, like response generation, tested independently before you rely on it.