AgentforceAdvancedaction-chainsorchestrationflowdesign-patterns
Chain dependent agent actions without fragile multi-step prompts
Real World Scenario
An insurance agent must verify policy, check coverage, and create a claim draft. When step two fails, partial claim records confuse adjusters and customers.
Expected Answer
• Model multi-step workflows as explicit state machine in Flow or Apex orchestrator, not LLM improvisation
• Each agent action returns structured status codes the agent interprets for next user message
• Use saga pattern with compensating actions rolling back prior steps on failure
• Persist workflow state in custom object linked to session ID surviving agent handoffs
• Limit LLM to conversational layer; business sequencing lives in deterministic code
• Define timeout and retry policies per action with customer-facing progress messages
• Integration test full chains including mid-chain failure and recovery paths
Follow-Up Questions & Answers
Click to expand — each follow-up includes a direct, interview-ready answer
Direct answer: Model multi-step workflows as explicit state machine in Flow or Apex orchestrator, not LLM improvisation Also consider: Each agent action returns structured status codes the agent interprets for next user message In practice: Use saga pattern with compensating actions rolling back prior steps on failure Optimize for scale and operational observability.
Architect Perspective
LLMs are poor workflow engines. Chain actions in code; let agents explain and confirm steps conversationally.