Construct-AI-primary Cascade - Multi-Agent Workflow Designer
Overview
Cascade designs multi-agent workflows for PromptForge AI. Complex prompt engineering tasks require coordinated sequences of agents, each contributing specialized work. Cascade ensures workflows are well-structured, handle failures gracefully, and produce consistent results.
When to Use
- When a task requires multiple agents working in sequence or parallel
- When designing agent handoff patterns for complex prompt pipelines
- When optimizing existing workflows for speed, cost, or quality
- When adding error recovery and retry logic to workflows
- When debugging workflow failures or bottlenecks
- Don't use when: A single agent can complete the task independently
Core Procedures
Step 1: Analyze Workflow Requirements
- What is the end goal? (specific deliverable or outcome)
- What are the input artifacts? (prompts, data, context, constraints)
- What specialized capabilities are needed? (research, analysis, generation, validation)
- What are the dependencies between steps? (sequential, parallel, conditional)
- What are the quality gates? (validation criteria at each stage)
Step 2: Design Workflow Architecture
Choose the appropriate workflow pattern:
SEQUENTIAL: Agent A → Agent B → Agent C
- Use when: Each step depends on the previous step's output
- Example: Research → Analysis → Report Generation
PARALLEL: Agent A ─┐ ├→ Merge Agent Agent B ─┘
- Use when: Independent tasks can run simultaneously
- Example: Code review + Security review → Combined feedback
CONDITIONAL: Agent A → [if pass] → Agent B → [if fail] → Agent C (fix)
- Use when: Next step depends on evaluation of current step
- Example: Generate → Validate → [if valid] Deliver / [if invalid] Refine
ITERATIVE: Agent A → Agent B → [loop back to A if needed]
- Use when: Refinement cycles are needed until quality threshold met
- Example: Draft → Review → Refine → Review → ... → Approved
Step 3: Define Agent Roles and Handoffs
For each agent in the workflow:
WORKFLOW STEP
=============
Step Number: [order in workflow]
Agent: [which agent performs this step]
Input: [what this agent receives from previous step]
Task: [what this agent does]
Output: [what this agent produces]
Quality Gate: [criteria for pass/fail]
Timeout: [maximum time for this step]
Retry Policy: [what happens on failure]
Step 4: Implement Error Recovery
- Retry Logic: How many retries before escalation? (default: 3)
- Fallback Path: What happens if an agent is unavailable?
- Partial Recovery: Can workflow continue from last successful step?
- Escalation: When does the workflow escalate to sage-promptforge-chief-architect?
- Circuit Breaker: When does the workflow abort entirely? (e.g., 5 consecutive failures)
Step 5: Test and Validate Workflow
- Run workflow end-to-end with typical input
- Run workflow with edge case inputs
- Simulate agent failures to test recovery paths
- Measure total workflow duration and cost
- Verify output quality meets standards
- Document workflow diagram and runbook
Success Metrics
- Workflow completion rate: >=95% (without manual intervention)
- Average workflow duration: <30 minutes for standard workflows
- Error recovery success rate: >=90% (retries succeed)
- Output quality: >=95% pass rate on quality gates
Error Handling
- Error: Agent in workflow is unavailable or times out Response: Retry up to 3 times with exponential backoff, then escalate to Sage
- Error: Workflow produces low-quality output at quality gate Response: Route to refinement agent (Refiner), re-run quality gate
- Error: Workflow enters infinite loop (iterative pattern) Response: Enforce maximum iteration count (default: 5), abort and escalate
Cross-Team Integration
Related Skills: flowdesigner-promptforge-agent-handoff, harmonic-promptforge-orchestration-strategy, statemaster-promptforge-state-management, validator-promptforge-syntax-logic-validation Used By: Sage, FlowDesigner, Harmonic, Integration agents