Salesforce hub — family 4, agent testing, evaluation and observability (2026-10-23)

Family 4 adds the agent testing reference (both runners, spec fields and the expectations they become, scorers, commands, limits, failures), a guide that tests the steward agent and gates CI on its results, and the observability reference (preview traces, sf agent trace, Session Tracing in Data 360, the OpenTelemetry export). Test spec blocks are now converted to metadata offline in the hub's checks.

markdownreferencev1NZZP-GBKSarea: changelogchangelogrelease notesagent testing familyobservability

Created 23 Oct 2026 · Modified 23 Oct 2026 · 1 version

Salesforce hub — family 4, agent testing and observability (2026-10-23)

Added

  • reference/agentforce/agent-testing.md: covers
  • the Testing Center (AiEvaluationDefinition) and Agentforce Studio (AiTestingDefinition) runners;
  • the spec fields and the expectations they become, and 11 scorers;
  • sf agent test and run-eval, and the Testing Connect API;
  • limits, exit codes and known failures.
  • guides/agentforce/test-an-agent.md: covers
  • a steward-agent spec with routing, action, outcome, multi-turn and guardrail cases;
  • creating and running it;
  • a TypeScript gate that fails CI on failed assertions, plus the CI step.
  • reference/agentforce/agent-observability.md: covers
  • preview traces with sf agent trace (dimensions, raw step types) and --apex-debug;
  • enhanced event logs;
  • Session Tracing objects in Data 360 and the beta OpenTelemetry export (72 hours).
  • work/tools/testspec2xml.mjs. check_code.py converts every # agent-test-spec YAML block to metadata with @salesforce/agents (what sf agent test create does) and packages it with sf project convert source.

Found while building

  • Failed assertions exit 0. Only test cases in ERROR exit 1.
  • Omitted expectedActions still asserts that no action ran ([]).
  • sf agent test results has no --use-most-recent, although its help says it does.
  • Salesforce's testing pages disagree on whether testing is sandbox-only.
  • Testing Center reports tool names (from reasoning.actions), not action declaration names, per Salesforce's sf-skills.

Still missing

  • No org was used, so result shapes are read from Salesforce's source.
  • Agentforce Studio testing is beta and its docs aren't out yet.
  • The Agent Health Monitoring API isn't verified.

Read this document as markdown — the same bytes an agent receives.