If you’ve been watching the Salesforce ecosystem over the past year, you’ve likely noticed that Agentforce is no longer a future consideration. It’s here, it’s actively being deployed, and 12,000-plus businesses are already running it in production. According to Salesforce’s own State of AI report, 83% of IT leaders expect AI agents to handle routine tasks independently within the next two years.
The good news is that building your first Agentforce agent is far more accessible than it sounds. The not-so-good news is that most guides either rush past the preparation stage or use outdated screenshots. This guide is updated for Spring ’26, accounts for the new Agentforce Builder (now generally available), and covers everything you need to go from zero to a working agent.
What Is Agentforce, Exactly?
Agentforce is Salesforce’s agentic AI layer, built directly inside your existing org. It is not a separate product you bolt on, it lives within the platform you already use.
The distinction that matters most: a traditional chatbot follows a fixed script. An Agentforce agent receives a trigger, reasons through the problem using Salesforce’s Atlas Reasoning Engine, and completes multi-step workflows without waiting to be told what to do at each stage. Think of the difference between a vending machine and a trained team member who understands context. You can build Sales Agents, Service Agents, Marketing Agents, Commerce Agents, and Platform Agents, making Agentforce relevant across virtually every business function.
The Four Building Blocks You Need to Understand First
Before you touch a single setting, it helps to understand what an agent is actually made of.
Subagents (previously called Topics, renamed in April 2026) define the specific domain your agent works within, such as Order Management or Lead Qualification. Each agent can have multiple subagents.
Instructions are written in plain, natural language and tell the agent how to behave, what tone to use, and what to avoid. The more precise these are, the more reliably your agent will perform.
Actions are the specific capabilities the agent can execute. These are built from Salesforce Flows, Apex classes, Prompt Templates, or MuleSoft API calls.
Guardrails set the hard operational boundaries. They define what the agent must never do, such as processing changes above a certain value or skipping a mandatory human handoff. All four sit beneath the Einstein Trust Layer, which handles data privacy and security at the platform level.
What You Need Before You Start
This section is where most tutorials let you down. Skipping prerequisites is the single biggest reason first-time builds fail.
Edition and licensing. Agentforce requires Enterprise, Unlimited, Performance, or Developer Edition. Salesforce uses a consumption model called Flex Credits, priced at approximately £1.60 per conversation at standard tier. If you are exploring before committing, Salesforce Foundations offers a no-cost pilot entry point with bundled credits.
Einstein Generative AI enabled. Go to Setup, search “Generative AI” in Quick Find, open Einstein Setup, and toggle Einstein on. This is the master switch. Nothing else works without it.
Data Cloud (Data 360). Technically optional, practically essential. Without Data Cloud, your agents answer from generic knowledge rather than your actual customer data. The difference in resolution quality is significant.
The right permissions. Builders need the Agentforce Admin, Data Cloud User, and Einstein Prompt Template Manager permission sets. Missing any one of these causes the setup to stall at a frustrating and opaque error message.
Clean CRM data. Your agent is only as good as the data it can access. Duplicate records and outdated information will undermine every response. If your org needs a tidy-up before you begin, our Salesforce Performance Optimisation service is a practical place to start. The Bulk Field Creator by 9to9Clouds can also help you configure the correct custom fields your agent will query, before you build.
A defined, narrow use case. Do not build an agent that handles everything. Choose one high-volume, low-complexity task to start: billing FAQs, case classification, or lead qualification all work well.
Building Your First Agentforce Agent: Step by Step
Step 1: Enable Einstein Generative AI
Setup, Quick Find, “Generative AI,” Einstein Setup, toggle ON. If the toggle is greyed out, your current edition does not support it. Confirm your licence with your Salesforce account executive before proceeding.
Step 2: Enable Agentforce
Setup, Quick Find, “Agents,” then open Agentforce Agents under Agent Studio. Toggle Agentforce on, then enable the Agentforce Default Agent. In Spring ’26, this screen is noticeably cleaner than before, with the new builder now fully available.
Step 3: Assign Permissions
Before anyone builds or tests, assign the Agentforce Admin, Data Cloud User, and Einstein Prompt Template Manager permission sets. This step is the most commonly missed, and skipping it causes a “Missing Permission Set” error that stops everything. Check every person on the team who will be involved.
Step 4: Open Agentforce Studio and Create a New Agent
Navigate to the Agentforce Studio app, click the Agents tab, and select New Agent. You can choose a pre-built template (the Service Agent template is the best starting point for a first build) or describe what you want to build to the AI Assistant, which will generate an initial configuration for you. For most first-time builders, the template route is faster and more reliable.
Step 5: Configure Your Subagents
Each subagent defines a specific domain of work. Give it a clear name, a plain-language description, and a scope statement. The Atlas Reasoning Engine reads these to determine which subagent to activate for a given customer input. Start with one or two subagents. In Canvas View, these appear as expandable blocks inside your agent script.
If you need support structuring complex, industry-specific subagents for telecoms, healthcare, or financial services, our Agentforce Development team can guide the configuration from the outset.
Step 6: Write Instructions and Add Actions
Instructions go directly into Canvas View in natural language. “Be helpful” is not an instruction. “Always summarise the open case before offering a resolution option” is. Write instructions that are specific and testable.
For actions, start with standard Flows rather than custom Apex for your first build. The Automation and Process Optimisation service covers the Flow architecture that powers agent actions most effectively. If your agent needs to interact with automation rules across your org, the Universal Automation Switcher makes toggling those rules on or off straightforward, without developer involvement.
Step 7: Set Guardrails and Test
Define what the agent must not do. For regulated industries, this step is non-negotiable. Then test inside Agent Builder Preview. Spring ’26 introduced Batch Testing, which allows you to run multiple test cases simultaneously using real interaction patterns, rather than testing one scenario at a time.
Step 8: Deploy to a Channel
Agentforce agents can go live on the web (via Experience Cloud), WhatsApp, SMS, Voice, and Slack. Deploy to your Sandbox first, validate with real users, then promote to production. Use Task Resolution Metrics (new in Spring ’26) to measure how effectively your agent is actually resolving interactions.
Common Mistakes to Avoid
Building an agent that handles too many scenarios at once is the most frequent error. Vague instructions are the second. Skipping the Data Cloud setup produces an agent that sounds knowledgeable but gives generic responses that frustrate customers rather than helping them. Always test in Sandbox. Always set guardrails before going live.
When to Build It Yourself vs. Bring in Support
If you have a clean org, a clear single use case, and Salesforce admin experience, a first build is manageable on your own using this guide. When the use case requires external integrations, industry-specific compliance, multi-agent orchestration, or a combination of OmniStudio and Agentforce, the complexity increases considerably.
At 9to9Clouds, we build Agentforce agents for businesses that need them to work reliably from day one. If you would like to talk through your use case before committing to a build, get in touch with our team. There is no obligation, and it may save you several weeks of iteration.
Frequently Asked Questions
Do I need Data Cloud to use Agentforce?
Technically no, but practically yes for any production deployment. Without it, agents cannot access real customer data and will give generic responses.
Can I build an Agentforce agent without coding?
Yes. The new Canvas View and the AI Assistant in Agentforce Builder are designed for admins working in natural language. Standard Flows and templates reduce the need for Apex significantly.
How much does Agentforce cost in 2026?
Standard consumption pricing is approximately £1.60 per conversation. Enterprise customers typically negotiate volume pricing. Salesforce Foundations offers a free pilot entry point.
What is Agent Script?
Agent Script is the new scripting language introduced in Spring ’26 that powers Agentforce Builder. It is human-readable and supports both natural language instructions and deterministic logic, giving you more control over how your agent reasons and transitions between subagents.
How long does it take to build a first agent?
A focused single-use-case agent with clean data and the right permissions in place can be built and tested in two to four weeks. More complex builds involving integrations or regulated workflows typically take six to eight weeks.