When I started teaching, I learned quickly that “teach the students” was not a useful instruction. Which students? Which subject? For how long, and toward what goal? A lesson only works once you decide what it is actually about. An Agentforce agent is no different. Before it can help anyone, it needs to know what it is for.
That clarity comes from two things: topics and instructions. They are quiet features, easy to skip past when you are excited to see your agent talk. But they are the difference between an agent that is genuinely helpful and one that wanders off and guesses.
What a topic actually is
A topic is a job area for your agent. It groups together a set of related things the agent can help with, along with the actions it is allowed to use to get them done.
Think of an agent that supports a sales team. You might give it one topic for “Order Status” and another for “Product Information.” Each topic is a defined lane. When a user asks something, the agent first decides which topic the question belongs to, then works inside that topic to respond.
This grouping matters because of how the planner works. The planner is the reasoning that picks which action to run. If your agent has fifty actions all thrown together with no structure, the planner has to consider all of them for every single question. Topics narrow the field. Once the agent recognizes that a question is about order status, it only looks at the actions inside that topic. Less noise, better choices.
A topic is not a feature you add for completeness — it is how you tell the agent which small, well-defined world a question lives in.
Instructions: the rules inside the lane
If a topic defines what the agent helps with, instructions define how it should behave while doing so. Instructions are written in plain natural language. No code, no syntax to memorize — just clear sentences describing what you expect.
Inside an “Order Status” topic, your instructions might say things like: always confirm the order number before looking anything up; if an order is delayed, apologize briefly and offer the expected ship date; never share another customer’s order details. These are the same kinds of guidance you would give a new team member on their first day.
Good instructions are specific and behavioral. “Be helpful” tells the agent nothing. “If the customer asks about a refund, explain our 30-day policy and then offer to start a return” tells it exactly what to do. The more precisely you describe the behavior you want, the more consistently the agent delivers it.
Why scope keeps an agent trustworthy
Here is the part beginners often miss. A narrow, well-scoped agent is a more reliable one, not a less capable one.
When a topic is clearly defined, the agent knows the edges of its own knowledge. It can recognize when a question falls outside its lanes and hand it off or politely decline, instead of inventing an answer. An agent with vague, overlapping topics has no clear sense of where it stops, so it tends to overreach. That is where unhelpful or wrong responses come from.
This is also why I encourage people to start small. One topic, a couple of clear actions, and tight instructions will outperform a sprawling agent every time. You can always add topics later as you learn what your users actually ask for.
The teaching parallel
I keep coming back to teaching because the mechanics are honestly the same. A good lesson plan has a defined subject — that is your topic. It has ground rules for how the class runs — those are your instructions. And within that frame, the student is trusted to think and respond. You do not script every word; you set the boundaries and the expectations, then let understanding fill the space.
An agent works the same way. You are not writing out every possible reply. You are defining the subject and the rules, then trusting the planner to reason within them. When the boundaries are clear, that trust is well placed.
Putting it together
A practical flow looks like this. You create a topic and give it a clear name and description so the agent can recognize when a question belongs to it. You attach the actions that topic needs — the things the agent is allowed to do within that area. Then you write instructions that shape its tone, its safeguards, and its step-by-step behavior.
The description of the topic is doing real work, by the way. The agent reads it to decide where a question fits, so write it the way you would explain the topic to a colleague: “Use this topic when the customer asks about the status, shipping, or delivery of an existing order.”
Start there, test how it responds, and refine the wording. Most of the improvement in an agent’s early life comes not from adding features but from sharpening these descriptions and instructions until the agent consistently understands what you meant.
Your next step
If you want the bigger picture of how agents reason and act, start with What Is Agentforce?. Then, to give your agent something concrete to do within a topic, walk through Building Your First Agentforce Action Without Writing Apex. You can also browse the full Agentforce category for more beginner guides.