triage-agentic-approachlisted

# Skill: triage-agentic-approach > **Invoked by:** `agentic-systems-architect` (primary). Also consulted by `agent-implementation-engineer` when a build reveals the task doesn't actually need the agentic machinery it was scoped with and should be re-triaged. > > **When to invoke:** "should we build an agent for this?"; "is this an agent or a workflow?"; "do we need multiple agents?"; "single agent with more tools, or a swarm?"; "which agent framework?"; any "how should we approach this agentically?" question. > > **Output:** a go/no-go verdict (defaulting to "a workflow or single call wins") + IF an agent survives: single-vs-multi + topology + framework + the cost/latency budget and guardrail level — every volatile framework/model fact dated and marked verify-at-use. ## Procedure 1. **Restate the task in the tree's terms.** Capture: the **goal** (what "done" means), whether the **steps are knowable in advance** (can you flowchart them?), whether the **step count is bounded**, what **tools/systems** it must touch and their **blast radius** (read-only vs writes to the real world), and the **cost/latency tolerance**. 2. **Run the agent-vs-not gate FIRST — this is the whole point.** The decisive question: **is the control flow knowable before runtime?** - **Yes, one shot** → a **single LLM call** with the right context (retrieval, examples, structured output). Stop here. - **Yes, multiple known steps** → a **workflow** (prompt chaining / routing / parallelization / orche