Support Agent Workbench
A production triage + reply agent with tools, guardrails, and a human-in-the-loop console. Includes a 42-case eval suite.
What it is
The Support Agent Workbench is a complete, opinionated foundation for a customer-support agent. It classifies incoming tickets, drafts grounded replies from your knowledge base, calls your tools (refunds, lookups, escalations), and routes anything risky to a human through a clean review console.
It's the architecture we wished existed — guardrails first, observability built in, and an eval suite so you can change the prompt without holding your breath.
What's inside
- ✓**🎯 Triage & routing** — intent + priority classification with a confidence threshold and fallback to humans.
- ✓**🛡️ Guardrails** — PII redaction, tool allow-lists, and refusal handling baked into the loop.
- ✓**🧰 Tool layer** — typed tool definitions for lookups, refunds and escalation — swap in your own.
- ✓**👤 HITL console** — a review UI where agents approve, edit or reject drafted replies.
What you get
support-agent/ agent/ — loop, prompts, tools console/ — HITL review UI evals/ — 42 test cases CLAUDE.md README.md .env.example
Eval report
Every template ships with a transparent eval run. Here's the latest.
Reviews
“Replaced three weeks of internal scaffolding. The HITL console alone is worth the price — our team trusts it because they can see every decision the agent makes before it fires.”
“The eval suite is the differentiator. I tweaked the routing prompt for our tone, ran the suite, and had confidence nothing regressed. That feedback loop would have taken me a week to build myself.”