Salesforce Agentforce — and ServiceNow, for contrast
The other four systems on this pack are things a company built for itself. This one is the pattern most
enterprises will actually deploy: a vendor platform with the intent classifier, reasoning loop, action
registry and trust layer already assembled. It is also the best-documented example of intent classification
in the whole pack — Agentforce's Topics are its classifier, and the mechanism is fully public.
6action types
5agent attributes
2×response relevance (pilot)
+33%end-to-end accuracy (pilot)
ReActnot chain-of-thought
A2A / MCPcross-agent protocols
Topics ARE the intent classifier
A Topic maps to one user intent or job to be done.
Classification is prompt-based against each topic's natural-language classification description, not a
trained classifier. In newer terminology each request maps to a subagent — “the AI's job
description for that moment” — which defines what actions are valid and which policies apply.
The topic is also the guardrail
Because a Topic bounds the action set the
planner may choose from, misclassification cannot produce an out-of-scope action. That structural
containment is more robust than prompt-level instructions, and it is the design idea worth stealing.
Vendor material, read accordingly
Unlike the Uber and LinkedIn pages, most
sources here are product documentation and marketing. Mechanics from developer.salesforce.com,
help.salesforce.com and the Salesforce Engineering blog are reliable; pilot performance claims
(2×, +33%) are vendor-reported and uncontrolled.
Full request pipeline — classification, reasoning, actions, trust
Read left to right for the request path; the bottom strip is the Einstein Trust Layer, which
wraps every model call rather than sitting at one point in the flow. The action block in the middle is
the answer to “how do agents connect to external systems” — six distinct mechanisms, from
no-code Flow to MuleSoft API to MCP.
Multi-step processes, record operations, business logic — no custom code
Prompt Template
Declarative / low-code
A grounded generation step, invoked as an action
Apex
Pro-code
Arbitrary custom logic on the platform
External Services
Pro-code / integration
Any OpenAPI-described REST API registered as an invocable action
MuleSoft API
Enterprise integration
Legacy and enterprise systems at scale. Topic Center defines Topics and Actions metadata at API design time, so existing API investment becomes agent-callable.
Agent Actions / MCP
Agent-to-agent
Other Agentforce agents, and MCP-exposed tools
Atlas: what it actually does
Three components define a generic agent:
State (short- and long-term memory), Flow (the logical framework guiding next steps), and
Side Effects (actions that impact the environment). On top of that, five attributes: Role, Data,
Actions, Guardrails, Channel. The loop is explicitly ReAct — reason, act, observe — until
the goal is met, with clarifying questions allowed mid-loop.
Why it scales
Salesforce's stated fix for early bottlenecks: replace synchronous,
tightly-coupled components with a publish–subscribe, event-driven graph of strongly-typed
function nodes, so tasks run concurrently and each module scales independently. Modules communicate through
“strongly typed well-publicized interfaces.”
Trust Layer, four mechanisms
Secure data retrieval — grounding uses
only data the executing user can access, respecting role-based controls and field-level security.
Dynamic grounding — assembled at run time, evolving with the conversation.
Data masking — sensitive values tokenised before leaving for an external LLM.
Audit trail — every prompt tracked through every step.
L3
Critical path — message to external API call to CRM update
ServiceNow — the same problem, a different centre of gravity
Worth putting side by side. Agentforce centres on one agent with a topic router;
ServiceNow centres on orchestrating many agents across a workflow engine. The Orchestrator is a
meta-agent — “a digital manager, assigning tasks, sequencing actions, and ensuring the right agent
executes the right task at the right time.”
