service-agentforce-human-escalation-configure

Configure Agentforce Agent → Human Escalation

Wire an Agentforce agent so it can hand a conversation off to a human agent through an Omni-Channel queue. This skill configures and verifies the escalation surfaces, then reports a single machine-readable JSON verdict. It is idempotent per developer name and refuses to write to production orgs. It runs standalone or as a stage in a larger setup flow (after the agent is created and active).

The agent, queue, outbound flow name, routed context object, service channel, and failure thresholds are all inputs, so the same skill serves any scenario (IT employee support, customer messaging, etc.) — the caller supplies the scenario-specific identifiers.

This skill is the authoritative owner of agent-to-human (outbound) escalation. service-agentforce-channel-configure may wire a basic outbound flow + connection block in the same pass as inbound channel setup, but it delegates full escalation configuration and verification here.

The honest boundary — what is deterministic vs. directive vs. runtime

Escalation in Agentforce is assembled from surfaces with different verifiability:

• Deterministic (this skill writes and/or round-trip verifies via the Metadata/Data API): canEscalate=true on the escalation topic (GenAiPlugin); outboundRouteConfigs on the agent's Messaging planner surface (GenAiPlannerBundle), where the SAME block couples outboundRouteName to outboundRouteType=OmniChannelFlow on a Messaging-class surface; a QueueBased outbound RoutingFlow with a non-null ActiveVersionId; a human queue with a QueueSobject for the context object, its QueueRoutingConfig bound to the queue, and at least one active human member (a queue with zero members can never route an escalation to a human, so it never reaches CONFIGURED); and the agent Active after republish.
• Directive / native policy (authored as Agent Script instructions, not a metadata counter): the default two consecutive failures threshold plus per-topic overrides (e.g. password reset = one failure). There is no metadata field that counts failed turns — the threshold is instruction text, verified by an ADK eval rubric, not a runtime read. See references/failure-policy.md.
• Runtime / manual only: the actual conversational trigger, AgentWork creation, and same-session context preservation. These cannot be driven headlessly — they are a documented manual test + eval rubric. See references/runtime-verification.md.