Globalcart Resolver Agent — Development skill for Claude Code

GlobalCart Operations Resolver Agent — Quest #4, Part A

A single autonomous agent that resolves GlobalCart customer support tickets (damaged items, late shipments, billing disputes) by investigating with real tools, applying the actual policy rules, and producing a structured decision — not a chatbot that guesses.

Why this needs to be an agent, not a fixed workflow

A ticket's real shape isn't knowable in advance — shipping issue, billing dispute, fraud risk, or a nonexistent order — so the investigation path can't be hardcoded. A fixed workflow (`order → profile → policy → decide`, always in that order) would waste calls on cases that resolve earlier, and can't skip steps it doesn't need (see scenario 6 below: the agent escalates after `check_return_policy` alone, correctly never calling `process_refund`). The model deciding which tools to call, in what order, and when it's done is what makes this an agent rather than a workflow, per Anthropic's [Building Effective Agents](https://www.anthropic.com/engineering/building-effective-agents).

Architecture

Built with **LangGraph** + **Claude** (`langchain-anthropic`), chosen over CrewAI or direct tool-use for one concrete reason: Stage 2 of this quest extends this into a multi-agent system, and LangGraph's node/edge graph model extends directly into multi-agent coordination without a rewrite. CrewAI is an awkward fit for a *single* agent (its whole value proposition is multi-agent crews), and building the tool-calling loop by hand would have skipped the actual framework-recognition skill this exercise is meant to build.

        ┌─────────┐
   ┌───▶│  agent  │◀──────────────────┐
   │    └────┬────┘                   │
   │         │ route()                │
   │    ┌────┴──────────────┬─────────┴──────────┐
   │    │                   │                     │
   │  real tool        submit_decision,       submit_decision,
   │  requested        <2 real calls done     >=2 done OR hit
   │    │