Multi-Step Workflow Orchestration

Advanced

Implement multi-step workflows with enforcement and handoff patterns · Difficulty 3/5

orchestrationmulti-stepdecompositionparallel

Prerequisites

Agentic Loop Lifecycle

Complex requests often contain multiple concerns that should be decomposed and handled efficiently.

Problem: Sequential Processing

For complex requests like "I've been charged twice, my discount didn't apply, and I want to cancel", agents may:

Process concerns one at a time (slow, many round-trips)

Fetch redundant customer data for each concern

Lose context between concerns

Solution: Decompose and Parallelize

Decompose: Break the request into distinct concerns

Share context: Fetch customer data once, share across all concerns

Investigate in parallel: Look into each concern simultaneously

Synthesize: Combine findings into a unified resolution

For Few-Shot Approach

If the agent handles single concerns well (e.g., 94% accuracy) but struggles with multi-concern requests (58%), the fix is adding few-shot examples demonstrating correct reasoning and tool sequences for multi-concern messages.

Key Takeaways

✓Decompose multi-concern requests and investigate in parallel
✓Share context (e.g., customer data) across parallel investigations
✓Few-shot examples improve multi-concern handling when single-concern works

Related Concepts

Multi-Step Workflow Enforcement

Programmatic prerequisites are more reliable than prompt-based ordering

Coordinator-Subagent (Orchestrator) Pattern

All communication flows through the coordinator (hub-and-spoke)