Agentforce Architecture Diagram

The components and their boundaries

1. Channel and execution identity

Establish who is asking and which identity each operation runs under. An employee assistant and a customer-facing service agent can have different access models. Document the Agent User boundary and test authenticated and anonymous paths separately.

2. Reasoning and available capabilities

The agent uses its configured subagents, instructions, and context to select work. Agent Script and deterministic logic can constrain the available paths. A natural-language instruction to avoid an operation does not replace a permission check or business validation.

3. Retrieval and business actions

Data Libraries support grounded answers through indexed content and retrieval. Flow, Apex, and API actions perform operations. A policy saying a refund is allowed is evidence for a decision; the refund action still needs its own eligibility checks, authorization, and any required human approval.

4. Outcomes and evidence

The agent returns an answer, reports an action result, or hands off. Define what must trigger escalation and what context the next person receives. Use session tracing to investigate a turn and a regression suite to check whether a fix holds across other requests.

Example: a customer asks for a refund

This is a proposed architecture walkthrough, not a report of a deployed customer system.

1. Identify the customer and order. Establish the permitted account scope before retrieving order details. A supplied order number alone is not proof of ownership.
2. Retrieve the applicable policy. Ground the answer in the current returns policy and its exceptions. Record which source supports the explanation.
3. Validate the operation. The refund action checks ownership, status, amount, and previous refunds. Prevent duplicate execution in the business layer.
4. Apply the approval gate. If the request exceeds a defined limit or needs an exception, route it to a person with the required authority.
5. Report the actual outcome. Confirm a refund only after the action reports success. Retain enough trace and audit evidence to reconstruct the decision without unnecessary customer data.

Questions for an architecture review

• Which identity reads each source and executes each action, and where is its access enforced?
• Which rules are deterministic, and which are only behavioral instructions?
• What happens when retrieval is empty, an integration fails, or an action is requested twice?
• Who can investigate the session, and what evidence survives long enough for an incident review?

Continue with soft and hard rule mapping, irreversible action gates, and the 21-check production readiness checklist.