Build a customer support agent

This cookbook builds a support agent that handles real production concerns: retrieval, customer context, review before send, trace inspection, and evals from resolved tickets.

Scenario

A customer opens a ticket. The agent searches product docs, checks account state, drafts a response, waits for human approval, and records the final outcome for future evaluation.

What you build

A support-triage workflow.
Tools for docs search and customer lookup.
A draft response step with structured output.
A human-review pause before sending.
A feedback path into an eval dataset.

Workflow shape

@workflow
async def support_agent_workflow(ctx: WorkflowContext, ticket_id: str) -> SupportOutcome:
    ticket = await ctx.step(load_ticket, ticket_id)
    account = await ctx.step(load_account_context, ticket.customer_id)
    docs = await ctx.step(search_support_docs, ticket.body)
    draft = await ctx.step(draft_support_reply, ticket, account, docs)

    decision = await ctx.wait_for_signal(
        "support_reply_review",
        timeout="3d",
        metadata={"ticket_id": ticket.id, "draft_id": draft.id},
    )

    if decision.status != "approved":
        return SupportOutcome(status="needs_changes", draft_id=draft.id)

    sent = await ctx.step(send_reply_once, ticket.id, draft.id)
    await ctx.step(record_support_eval_case, ticket.id, draft.id, sent.id)
    return SupportOutcome(status="sent", message_id=sent.id)

The agent is useful because every step is visible and recoverable.

Trace review

For each support run, reviewers should see:

ticket input,
retrieved docs,
account context used,
draft response,
approval decision,
final send receipt.

This makes support QA concrete. A bad answer can be traced to a retrieval miss, an account-state error, or a prompt failure.

Eval loop

When a reviewer edits the draft, record the corrected response as an eval case. Later prompt and model changes can replay the same ticket and compare against the approved answer.

Production checks

Tenant and deployment IDs are included on every direct HTTP call.
Retrieved docs are stored as trace evidence.
The send step uses an idempotency key.
Rejected drafts stop before the send step.
Reviewer edits can become eval cases.