Framework Cookbook: Support Copilot

The reference application is a support copilot for customer-service teams. It does four things:

• retrieves relevant policy and account context
• drafts a support response
• proposes next actions such as refund, escalation, or follow-up
• requires human approval before high-risk side effects

This workflow is deliberately representative: it combines retrieval, tool use, structured outputs, and approval gates without being tied to one vendor.

With the OpenAI Agents SDK, the support copilot maps naturally to an agent with typed tools, explicit instructions, and a runtime that handles the action loop for you.

agent = Agent(
  name="SupportCopilot",
  instructions="Draft support responses, cite policy, and escalate risky actions.",
  tools=[lookup_account, search_policy, draft_refund_plan],
)

This approach is strong when you want a direct agent abstraction with clean tool orchestration and minimal graph plumbing.

In LangGraph, the same workflow is easier to model as explicit state transitions: classify request, retrieve context, draft response, validate, then route either to approval or final output.

START -> classify -> retrieve -> draft -> validate -> {approve | finalize}

This is a good fit when the workflow has branching, recovery logic, or durability requirements that should be visible in the architecture.

PydanticAI makes the support copilot attractive when typed inputs and outputs are central. Its value shows up in structured result validation, safer tool signatures, and explicit data models.

Vercel AI SDK becomes compelling when the support copilot is tightly tied to a web UI: streaming drafts, message-based interactions, and attachment-aware chat surfaces are first-class concerns.

The key lesson is that the workflow stays the same. What changes is where the framework gives you leverage: runtime orchestration, type safety, or UI delivery.

When comparing framework implementations, do not ask only “which one looks cleaner?” Compare the parts that affect production work:

• state model and workflow visibility
• tool definition ergonomics
• typed validation support
• human-approval integration
• streaming and UI fit
• testing and observability hooks

A cookbook is valuable when it keeps the business workflow constant and changes only the implementation surface. That makes framework tradeoffs visible instead of hypothetical.