What Agentforce 2.0 Actually Is
Agentforce is Salesforce’s platform for building autonomous AI agents that operate inside your CRM. Unlike a chatbot that follows scripted flows, an Agentforce agent reasons over your data, picks tools on its own, and takes real actions — creating cases, updating opportunities, drafting emails, or escalating to humans when confidence drops.
Version 2.0 (announced at Dreamforce and expanded in the Spring ‘26 release) shipped deeper reasoning, richer action libraries, and tighter Data Cloud integration. If you evaluated Agentforce 1.0 and found it thin, 2.0 is the release worth a second look.
The Four Building Blocks
Every Agentforce agent is made of the same four pieces. Understand these and the rest of the platform clicks into place.
1. Topics
A topic is a bucket of related jobs the agent is allowed to handle — “order management,” “account renewals,” “lead qualification.” Each topic has a scope (what it can do), instructions (how it should behave), and a set of actions it can call.
Topics are how you prevent agent sprawl. An agent scoped to “order management” should not suddenly answer HR questions.
2. Actions
Actions are the concrete things an agent can do. Out of the box you get standard actions like “Create Case,” “Draft Email,” and “Query Records.” You can also wrap any Apex class, Flow, or prompt template as an action.
Actions are where almost all the real work of building an agent lives. A well-designed action library makes the agent useful; a sloppy one makes it unreliable.
3. The Atlas Reasoning Engine
Atlas is what sits between the user’s question and the agent’s response. It classifies intent, picks the right topic, plans the steps, chooses actions, and verifies the output before returning it. In 2.0, Atlas runs a multi-step reasoning loop — plan, act, observe, revise — instead of a single-shot classification.
The practical consequence: agents handle ambiguous or multi-step requests far better than 1.0. The tradeoff: every reasoning step consumes Einstein request credits, so you need to monitor usage.
4. The Einstein Trust Layer
The Trust Layer is Salesforce’s guardrail stack. It masks PII before prompts leave your org, filters toxic or biased output, enforces zero-retention with the underlying LLM provider, and logs every interaction for audit.
Trust Layer is on by default. You should not disable it.
How to Deploy Your First Agent
Here is the minimum viable rollout that won’t embarrass you in production.
Step 1: Pick a Narrow Use Case
Resist the temptation to build a “do everything” agent. Pick one high-volume, low-risk job — answering “where is my order,” drafting follow-up emails, summarizing a case history. Narrow scope makes evaluation tractable.
Step 2: Inventory the Data
The agent is only as good as what it can see. List every object, field, and external data source it needs. Map Data Cloud connections if you’re pulling from a warehouse.
Step 3: Build the Actions First
Before wiring up the agent, build each action as a Flow or invocable Apex method. Test each one independently. An agent that calls broken actions looks broken — even when the reasoning is correct.
Step 4: Write Topic Instructions Like Prompts
Topic instructions are natural-language prompts. Be specific. “When the user asks about an order, call Get_Order_Status first, then confirm shipping address before creating a return” works better than “help with orders.”
Step 5: Evaluate with a Test Set
Before going live, run 50–100 representative inputs through the agent and grade the responses. Agent Builder includes a testing surface for this. Do not skip this step.
Agentforce vs. Einstein Copilot
A common source of confusion: Einstein Copilot was the 2024 branding for the in-app AI assistant. Agentforce is the broader platform that Copilot is now a part of. If you see “Einstein Copilot” documentation, treat it as an earlier name for the same thing — Salesforce has consolidated the naming.
What It Costs
Agentforce is licensed by “conversation” (one interaction cycle) rather than by user. As of the 2026 pricing updates, conversations are priced per 1,000 blocks, with tiered volume discounts. Einstein request credits (consumed by Atlas reasoning) are a separate meter.
Budget at least a small pilot SKU before proposing a broad rollout — usage-based pricing surprises finance teams that are used to per-user models.
Common Pitfalls
Over-scoping topics. A single topic handling eight unrelated jobs is a debugging nightmare. Break it up.
Under-specifying actions. If an action’s input/output schema is vague, the agent will pass wrong arguments. Document every parameter.
Skipping Trust Layer config. PII masking rules are org-specific. The defaults are reasonable but not sufficient for regulated industries — tune them.
No human escalation path. Every agent needs a graceful handoff to a human when confidence is low. Build it in from day one.
When Not to Use Agentforce
Agentforce is overkill for deterministic workflows. If the job can be described as “when X happens, always do Y,” a Flow is cheaper, faster, and more predictable. Reserve agents for jobs that genuinely require reasoning over ambiguous input.
Frequently Asked Questions
Does Agentforce replace Flow?
No. Flow is still the right tool for deterministic automation. Agentforce is for jobs that require reasoning.
Can I use my own LLM?
Salesforce currently hosts the models through its own trust layer. Bring-your-own-model is available for specific regulated scenarios via BYOLLM partnerships, not as a general option.
Is Agentforce available in every edition?
Agentforce requires Enterprise Edition or higher plus an Agentforce SKU. Developer Edition orgs get a limited trial allocation.