Salesforce

Agents act on the customer's systems through documented actions: Agentforce Builder connects agents to data, channels and existing Salesforce assets including Flows, Apex, MuleSoft APIs and prompt templates, and out-of-the-box actions cover common jobs such as order and reservation management.

Sourcesalesforce.com/agentforce/agent-builderread 2026-09-18

Agent Script defines how work moves through an agent: every request begins at an agent router that sets initial variables and classifies the request, then the agent processes a subagent and transitions between subagents along paths the script defines. Across agents, an orchestrator agent can invoke specialized agents in other Salesforce orgs and outside Salesforce through the Agent API, and a flow action can invoke an agent as one step of a longer automation.

Sourcedeveloper.salesforce.com/docs/ai/agentforce/guide/ascript-flow.mdread 2026-09-18

Agents are grounded in the customer's own material. An Agentforce Data Library connects an agent to trusted unstructured or semi-structured sources, such as files or a knowledge base, which the customer adds to the library and the agent retrieves from when it answers, so new material enters by adding it rather than by retraining.

Sourcedeveloper.salesforce.com/docs/ai/agentforce/guide/headless-examples-agent-adl.mdread 2026-09-18

Escalation to a person is a built-in agent utility: utils.escalate hands the conversation to a human service rep through an Omni-Channel routing connection the customer defines, and instructions tell the agent when to call it. Agent Script structures what the agent may do, with every request passing an agent router that classifies it before any subagent runs and explicit transitions between subagents, and Salesforce's own guidance recommends a person check model responses before they are shared in most use cases.

Sourcedeveloper.salesforce.com/docs/ai/agentforce/guide/ascript-ref-utils.mdread 2026-09-18

Salesforce's compliance site publishes attestations scoped to the agent layer itself: a SOC 2 report and a SOC 3 report for the Einstein Platform and Agentforce on Hyperforce, an ISO 42001 certificate covering the Salesforce AI Platform and Agentforce, a PCI attestation of compliance for Data Cloud, the Einstein Platform and Agentforce, a C5 report, an IRAP assessment and Spain ENS High for the Einstein Platform and Agentforce, alongside FedRAMP High and Moderate and ISO 27001 categories.

Sourcecompliance.salesforce.com/enread 2026-09-18

Every step of an agent interaction is recorded: the Agentforce Session Tracing data model and the Generative AI Audit and Feedback data model capture turns, messages, LLM calls, actions, metric scores and feedback, and the Session Trace OTel API exports a full session as OpenTelemetry-format JSON for any observability tool or OTLP collector. Trace files can also be managed from the command line during development, and the Einstein Trust Layer keeps an audit trail of generations.

Sourcedeveloper.salesforce.com/docs/ai/agentforce/guide/otel-api.mdread 2026-09-18

Agent Script variables let an agent deterministically remember information across conversation turns and keep context throughout the session, shared by every subagent, and a session ID carries that context across invocations. Data 360 and Customer 360 give agents live customer records, but those are the business records themselves, not a memory the agent keeps.

Sourcedeveloper.salesforce.com/docs/ai/agentforce/guide/ascript-ref-variables.mdread 2026-09-18

Salesforce runs on Hyperforce, its public-cloud infrastructure, where customer data is generally stored in the country where the customer's org is located, provided the services in use are available in that Hyperforce country, and a secure backup of an org can be held in a secondary Hyperforce region.

Sourcesalesforce.com/platform/public-cloud-infrastructureread 2026-09-18

Salesforce ships customer-adoptable packaged assets: Agentforce Builder starts agents from out-of-the-box subagents, actions and templates that cover common jobs such as order management, reservation management and general FAQs, which the customer then extends and refines with natural-language instructions.

Sourcesalesforce.com/agentforce/agent-builderread 2026-09-18

Work reaches Agentforce agents without a person asking. Salesforce documents headless agents that are called without a human trigger from a user interface: an Apex trigger or a scheduled job calls the agent as part of a background process, and a standard AI Agent action lets any flow, including one started by a record change, invoke an active agent with a prompt.

Sourcedeveloper.salesforce.com/blogs/2025/04/build-headless-agents-with-the-agent-api.htmlread 2026-09-18

The customer chooses the model. An administrator sets the org-level model for Agentforce on the Agentforce Agents page in Setup, and Agent Script overrides it for a specific agent or subagent with a model configuration naming any supported model. The supported list spans Salesforce-managed models from OpenAI, Anthropic on Amazon Bedrock, Google Gemini on Vertex AI and Amazon Nova, most running inside the Salesforce trust boundary, and customers can also bring their own model.

Sourcedeveloper.salesforce.com/docs/ai/agentforce/guide/ascript-model.mdread 2026-09-18

Agents are callable from outside Salesforce. The Agent API opens a conversation with an agent, sends prompts synchronously or as a server-sent event stream with progress, text-chunk and end-of-turn events, and closes the session, authenticated through a connected app with OAuth scopes, with a published Postman collection and Node.js client. Inside the platform, an invocable action runs any active agent from Apex or Flow, and Salesforce Hosted MCP Servers expose Salesforce data and automation to external AI agents and LLM applications over MCP.

Sourcedeveloper.salesforce.com/blogs/2025/04/build-headless-agents-with-the-agent-api.htmlread 2026-09-18

Agents are tested before release with scored test cases. A test spec describes test cases as utterances with the expected agent behavior, can set context variables and include out-of-the-box metrics in the output, and is created and run in a development org from the command line or the Agentforce Testing Center, with pass and fail results.

Sourcedeveloper.salesforce.com/docs/ai/agentforce/guide/agent-dx-test.mdread 2026-09-18

Nothing on the Agentforce estate documents an agent operating a browser or desktop interface. Agents act through actions: flows, Apex, APIs, prompt templates and MCP tools, reached through the agent router and subagents in Agent Script.

Sourcedeveloper.salesforce.com/docs/llms-genai.txtread 2026-09-18