Want Smarter AI Agents? It All Starts With Better Instructions in Agentforce

Agentforce is smart, because it doesn’t just follow orders, it reasons through them.

Instead of relying on rigid, hard-coded logic to accomplish specific goals, Agentforce leverages the Atlas Reasoning Engine to intelligently orchestrate actions.

As a developer, you design those actions using flows, Apex, or prompt templates and then guide the engine using natural language instructions. This allows your agent to act more intelligently and flexibly in real-world scenarios.

Think of it like cooking: even the best ingredients (Apex, flows, templates) won’t lead to a great meal without a good recipe. Instructions are the recipe that tells Agentforce how to combine everything just right.

To get the best results, you’ll need to craft different types of instructions: topic, action, and input/output. These help bridge the gap between your agent’s reasoning and its ability to perform meaningful actions on a user’s behalf.

In this blog post, we’ll walk you through how to write clear, effective natural language instructions to build smarter, more capable agents with Agentforce.

How to Write Effective Instructions for Agentforce AI Agents

To steer your agents in the right direction, you need clear, organized instructions that match user queries, task execution, business goals, team responsibilities, and workflows.

Here’s how you can get your AI agents working efficiently by crafting smart, effective instructions.

Topics

Topics act as organized containers that guide the agent on what tasks it can perform. They help make sure the agent answers questions accurately and consistently. Salesforce offers ready-made standard topics that you can start with and customize, or you can build your own topics from scratch.

Key components of a Topic setup:

• Classification: This is a brief summary of what the topic covers. It helps match the topic to the user’s question.
  Example for Order Management: “Handles customer questions about order status and returns.”

• Scope: This defines exactly what the agent should do and what it should avoid.
  Example: “Resend reservation confirmations only; do not create new reservations.”

• Instructions: These are clear, practical rules that the agent follows within the topic. Instructions shouldn’t be strict business rules but helpful guidance.

  Use positive language for clarity — saying “always” is better than “don’t,” because it makes the expected behavior clear. Also, combine related instructions in one place for easier understanding.

  For example:

  “Always ask if the device is iOS or Android before troubleshooting.”
  “Never ask for an ID; use the record name instead.”
  “If no order number is provided, escalate the case to a live agent.”

Topic Classification Description

Classification describes what your topic does. It’s used to direct the user to the appropriate conversation flow. Think of it as a short summary of what the topic is about.

Scope defines what the agent is allowed to do within that topic. It’s like a job description – be clear about what the agent can and can’t handle once it’s in that topic.

Action Instructions
Action instructions guide the agent on what an action does and when to use it. Writing them clearly helps the agent respond correctly and stay consistent.

Tips for writing Action instructions:

• Be Clear and Brief
  Use 1–3 simple sentences to explain what the action is for, what it does, and when to use it.

  Before: “Updates a status field.”
  After: “Changes the Task status to ‘Completed.’ If the Task doesn’t exist, create a new one with the given status.”

• Use Distinct Action Names
  Avoid using similar names for different actions – it can confuse the agent. Use varied but related verbs to help it distinguish between actions.

  Instead of: “Get Project Details” and “Get Task Details”
  Use: “Locate Project Details” and “Retrieve Task Information”

• Mention Dependent Actions
  If an action depends on another one, make that clear so the agent knows the right order.

  Example: “Run IdentifyCustomerAccount before RetrieveBillingHistory to get the Account ID.”

Inputs and Outputs Instructions

Make sure you clearly define:

• Inputs: What information the action needs, and what format it should be in.
  Example: accountId – The 18-digit Account ID, either from a previous step or entered by the user.

• Outputs: What the action returns.
  Examples:
  balance – The current account balance in numeric form.
  taskSummary – A short description of what was completed, like “Resolved payment issue.”

Test and Improve

Once your instructions are written:

• Test the agent using tools like Agent Builder and Plan Tracer to see how well it follows your instructions.
• Refine your language if needed, adjust your prompts or wording until the agent responds the way you want, every time.

Grazitti’s Custom Solutions That Amplify Agentforce Performance

Agentforce packs a punch.

But its real power shines when it’s molded to fit your business like a glove, thinking, responding, and adapting to your unique rhythm.

Customization is the game-changer that turns a strong platform into a strategic ally. Without it, you’re only scratching the surface of what Agentforce can truly do, like taking a high-performance car off-road instead of letting it loose on the track.

At Grazitti, we’ve spent over ten years mastering the art of customizing Salesforce to align with businesses’ evolving needs.

We understand that in today’s dynamic landscape, cookie-cutter solutions just don’t cut it. Every business faces different challenges, and those deserve tailored answers.

As Salesforce continues to innovate, we’ve been at the forefront, engineering custom solutions that enhance Salesforce and redefine how Agentforce delivers value.

Our tailored enhancements allow Agentforce to grow with your business, driving smarter decisions, streamlined operations, and a customer experience built to reflect your brand’s individuality.

At Grazitti, customization is the catalyst that unleashes Agentforce’s full potential and powers transformative business outcomes.

We’ve helped businesses like yours to optimize the potential of Agentforce.

Conclusion

Agentforce offers a new paradigm for businesses.

Instead of relying on traditional programming constructs, you guide a large language model using clear, high-level instructions. These instructions act as the Agent’s internal compass, teaching it how to reason through tasks, chain actions, and adapt to dynamic conditions.

It’s a shift in mindset – from thinking about how to code, to thinking about how to guide. And it’s this shift that unlocks the full potential of Agentforce.

To utilize it’s full potential, follow these best practices:

• Start small: Don’t aim for perfection on day one. Begin with a focused use case and use it to understand the core building blocks.
• Be clear: Each topic should serve a single purpose. Avoid blending goals across Classification, Scope, and Instructions.
• Use strong verbs: In both Scope and Instructions, use commanding, action-oriented language to steer the Agent with precision.
• Don’t over-label: If your Scope and Instructions are well-crafted, Example inputs may not even be necessary.
• Experiment intentionally: Try out different instruction formats and observe how they influence output. Learn, refine, and repeat.
• Tap into the community: Draw insights from other ISVs and admins. Real-world experience is often the fastest way to troubleshoot and improve.

Because at the end of the day: great agents aren’t coded, they’re guided.