The Beta is genuinely useful for routine work and genuinely incomplete. The rule that pays off in practice is: use the agent for the routine 80 percent, keep manual Setup skills sharp for the other 20 percent. Treat every Apply click as a real metadata change that ships through your normal deployment process, including sandbox first.
- Enable Setup with Agentforce on a sandbox first
Setup, Feature Settings, Setup with Agentforce, Enable. Do this in a sandbox so the team can try requests without risk. Production rollout happens after the team is comfortable.
- Try low-risk requests to calibrate expectations
Ask the agent to create a custom field, build a simple validation rule, add a value to a picklist. Observe the preview quality and the agent reasoning. This calibrates what to trust without risk.
- Read every preview before clicking Apply
The preview is the safety mechanism. Reading it carefully catches the edge cases the agent missed (wrong picklist value, wrong page layout, wrong dependent field). Apply without reading is the failure mode that produces incidents.
- Use the agent for multi-step routine work
Requests like "create a Tier field on Account and assign it to the Sales Manager permission set" save the most time. Single-field changes are fast either way; the multi-step requests are where the agent earns its keep.
- Keep manual Setup skills sharp
The Beta does not cover every Setup operation, and complex debugging still requires hands-on UI work. Continue training admins on the underlying Setup tree, not just on agent requests.
- Roll changes to production through your normal deploy path
An agent-generated metadata change is still a metadata change. Push it from sandbox to production through your change set, source-tracked sandbox, or DevOps Center pipeline. The agent is the editor, not the deploy mechanism.
- Track Beta limitations against your roadmap
Watch the release notes for each Salesforce release. Action coverage expands per release, and capabilities that were missing in the Spring may ship in the Summer.
Which orgs (sandbox, production, scratch) have the feature flag on. Default off in production until the admin team is comfortable.
Which permission sets grant access to the Setup chat panel. Defaults to System Administrator; can be relaxed for delegated admins.
Which Setup operations the agent supports in the current release. Tracked in release notes; expands per release.
How much detail the preview shows for dependent changes (default vs verbose). Verbose helps admins still calibrating trust.
Whether AgentInteraction records cross-link with SetupAuditTrail entries for unified compliance review. Enabled by default once the feature is on.
- The Beta does not cover every Setup operation. Action coverage is documented per release; check the release notes before assuming a specific capability is supported.
- The preview is the only safety mechanism between the agent and real metadata. Apply without reading is the most reliable way to produce an incident.
- The agent runs as the calling admin. It does not grant permissions the admin does not have, but it also does not warn when the admin is using their broad permissions on a sensitive object.
- Multi-step refinement only works inside an uncommitted preview. Once Apply is clicked, further changes require a fresh request and a fresh preview.
- Agent-generated changes still need to flow through your normal deploy pipeline. The agent is the editor; the deploy mechanism is unchanged.