Handing over control to a custom AI Agent requires verifying data readiness, permission scoping, and testing first. Without those checks, an agent can act on incomplete information or exceed intended boundaries.
What data and permission checks come before any build?
Start with an inventory of the objects, fields, and knowledge sources the agent will touch. Confirm that Grounding sources such as Salesforce objects, flows, and Data 360 (formerly Data Cloud) are complete and accurate before writing the agent definition. An agent asked to surface order history may fail or return a partial answer if a required field was never added to its scope.
Give the agent its own named user, with permission sets that grant only the object and field access its topics need. Do not grant View All or Modify All, and let the agent inherit sharing rules rather than bypass them. The same user needs a clear record of which flows and Apex classes it may invoke. Keep the permission set and agent definition in the same change set so mismatches show up in review.
Map how customer context will be verified on every interaction. Service agents commonly use a verified-customer filter so the agent cannot pull records outside the authenticated user’s scope. Document those filters with the agent definition, and do not rely on a single field that can be null.
How do you keep the agent’s job narrow enough to succeed?
Define one concrete job and refuse anything broader. An example agent that qualifies inbound leads can also update lead status and create follow-up tasks, but it should not also handle case creation or inventory checks.
Use the Agentforce Builder’s assisted creation flow to generate an initial set of subagents and actions from a short role description, then prune anything outside that job. Review the generated system messages and Agent Script steps, and keep those steps explicit rather than relying only on generative reasoning. Remove unintended actions before the first test run, and tighten instructions if later tests show the agent stepping outside its job.
What security and governance questions should you settle first?
Classify every action as public or private. Public actions can be called from more channels; private actions need additional verification or should stay in internal experiences. External tools registered through MuleSoft or MCP servers should be allowlisted individually, with strict quotas so one integration cannot exhaust Flex Credits.
Require that every custom Apex or Flow action runs in user mode with sharing checks. Do not pass raw record IDs from user input into those actions; resolve the ID through the agent’s verified context first. Review Trust Layer logging, including where events are stored. Data masking in the Trust Layer is currently disabled for Agentforce, so permission scoping and grounding choices remain the primary controls over what a prompt can include.
Establish an allowlist of topics and actions the agent may invoke and a separate list of actions that require human approval. Store both with the agent definition and change them through the same process you use for Apex or Flow updates. Many teams also export the agent configuration as metadata before each production update so they can restore a known-good state.
What should a test plan cover before handover?
Build a test plan that includes happy-path conversations and deliberate misuse attempts, such as incomplete customer data, conflicting instructions, and requests that try to expand the agent’s scope. Document pass criteria before testing begins.
Use the Agentforce Testing Center and the preview tools inside Builder to run these scenarios with mock data first. Capture traces so you can see which actions fired, and record the instruction change that corrected any failure. Previews and some test runs can consume credits, so include that usage in the post-launch monitoring plan. Route a controlled set of real channel traffic through the agent before expanding access, and name who can disable or revert the agent if needed.
Who owns monitoring after the build is handed over?
Assign a single owner for consumption monitoring before the agent goes live. That owner reviews Flex Credit usage and conversation counts on a set cadence and can pause the agent if spend exceeds an agreed threshold. Building an agent in Builder is generally not metered; production actions are, and previews and some test runs can also consume credits.
The owner should know how to disable specific actions or roll back to a prior version. Ongoing maintenance includes reviewing Agent Optimizer suggestions and updating grounding sources when objects or knowledge articles change.
A custom build outside Agentforce is the more practical choice when the workflow spans multiple non-Salesforce systems with heavy model-portability needs, or when the team needs direct control over the underlying model.
Frequently asked questions
How many agents should share a single agent user?
Create one dedicated agent user per agent. Separate users make it possible to apply distinct permission sets and simplify rollback when only one agent needs to be paused.
What happens if an agent attempts an action outside its defined topics?
The platform can block the action when topics and actions are explicitly restricted. Monitor traces anyway, because blocked attempts can mean misuse or an instruction set that needs tightening.
Can external systems call Agentforce agents directly?
Agents are typically exposed through supported channels such as chat, voice, Slack, or Experience Cloud portals. Direct external access generally requires a registered path such as an MCP server or a MuleSoft API, and it still runs under the scoped agent user permissions.
How often should grounding sources be refreshed?
Review grounding sources whenever the underlying objects, knowledge articles, or Data 360 segments change. Stale data produces incorrect or outdated responses even when the agent logic itself remains sound.
Who owns the test artifacts after the build is complete?
The team that will operate the agent should retain the test plan, traces, and pass criteria. Those artifacts become the baseline for later Agent Optimizer work and for any incident investigations.
Agentforce & AI Agents
Autonomous agents that predict, automate, and execute revenue-driving actions across sales, service, and marketing.
Read Next

Agentforce Governance and Security: Built-In Controls That Actually Work
Agentforce agents inherit Salesforce security and add dedicated controls for autonomous use. Learn how the Trust Layer, execution contexts, Gateway policies, and observability keep agents safe without extra tools.

Verifiable Evidence in Agentforce Shows Exactly How Agents Reason and Act
Agentforce records every reasoning step, tool selection, variable change, and outcome in structured traces that you can open and review directly in the platform.

Data Honesty Keeps Agentforce Agents Accurate and Reliable
Data honesty requires complete, consistent, and governed Salesforce records so Agentforce agents avoid errors and deliver trustworthy answers. Learn the practical steps to prepare your data before deployment.

