← ClaudeAtlas

agentic-workflow-designlisted

Use when the user wants to design an AI agent, multi-step LLM workflow, tool-using assistant, or automation with LLM calls — or is deciding whether they need an "agent" at all. Picks the simplest workable pattern, defines tools, control flow, guardrails, and failure handling before any code.
Krosebrook/augmentor-packs · ★ 0 · AI & Automation · score 72
Install: claude install-skill Krosebrook/augmentor-packs
# Agentic Workflow Design Default opinion: **use the simplest thing that works.** A single well-prompted call beats a chain; a fixed chain beats an autonomous agent. Add autonomy only when the steps genuinely can't be known in advance. ## Pattern ladder (choose the lowest rung that meets the need) 1. **Single call** with good context and output schema. 2. **Prompt chain:** fixed sequence of calls, each with a narrow job; code checks between steps. 3. **Router:** classify input, send to a specialized prompt/tool path. 4. **Parallelization:** independent sub-tasks run concurrently, then aggregate (or multiple attempts + vote). 5. **Orchestrator–workers:** a planner call decides sub-tasks dynamically; workers execute. 6. **Evaluator–optimizer loop:** generate → critique against criteria → revise, with a max-iteration cap. 7. **Autonomous agent:** model chooses tools in a loop until done. Requires strong tools, sandboxing, stopping conditions, and observability. ## Design worksheet (fill this before code) - **Job:** input → output, with 3 concrete examples. - **Why not one rung lower?** (Required answer.) - **Steps / control flow** diagram in text. Which decisions are made by code vs. by the model? - **Tools:** name, one-sentence purpose, input schema, output, side effects. Prefer few, well-described, hard-to-misuse tools. Read-only by default. - **State & memory:** what's passed between steps; what persists. - **Stopping conditions:** success criteria, max steps/iterations, b