设计生产级多Agent工作流,明确选择顺序、并行或分层模式,定义交接契约、故障处理与成本/上下文控制。适用于架构多步骤Agent流水线、在单Agent与多Agent方案间选型,…
下载这个 Skill 需要先登录(免费注册,只需用户名和密码)。去登录 →
按元信息、示例、可执行性、完整性等 13 项自动检查,供作者自查改进;不影响用户评分。
# Agent Workflow Designer **Tier:** POWERFUL **Category:** Engineering **Domain:** Multi-Agent Systems / AI Orchestration --- ## Overview Design production-grade multi-agent workflows with clear pattern choice, handoff contracts, failure handling, and cost/context controls. ## Core Capabilities - Workflow pattern selection for multi-step agent systems - Skeleton config generation for fast workflow bootstrapping - Context and cost discipline across long-running flows - Error recovery and retry strategy scaffolding - Documentation pointers for operational pattern tradeoffs --- ## When to Use - A single prompt is insufficient for task complexity - You need specialist agents with explicit boundaries - You want deterministic workflow structure before implementation - You need validation loops for quality or safety gates --- ## Quick Start ```bash # Generate a sequential workflow skeleton python3 scripts/workflow_scaffolder.py sequential --name content-pipeline # Generate an orchestrator workflow and save it python3 scripts/workflow_scaffolder.py orchestrator --name incident-triage --output workflows/incident-triage.json ``` --- ## Pattern Map - `sequential`: strict step-by-step dependency chain - `parallel`: fan-out/fan-in for independent subtasks - `router`: dispatch by intent/type with fallback - `orchestrator`: planner coordinates specialists with dependencies - `evaluator`: generator + quality gate loop Detailed templates: `references/workflow-patterns.md` --- ## Recommended Workflow 1. Select pattern based on dependency shape and risk profile. 2. Scaffold config via `scripts/workflow_scaffolder.py`. 3. Define handoff contract fields for every edge. 4. Add retry/timeouts and output validation gates. 5. Dry-run with small context budgets before scaling. --- ## Common Pitfalls - Over-orchestrating tasks solvable by one well-structured prompt - Missing timeout/retry policies for external-model calls - Passing full upstream context instead of targeted artifacts - Ignoring per-step cost accumulation ## Best Practices 1. Start with the smallest pattern that can satisfy requirements. 2. Keep handoff payloads explicit and bounded. 3. Validate intermediate outputs before fan-in synthesis. 4. Enforce budget and timeout limits in every step.