New video up. HR onboarding agent.
Full workflow:
→ Reads passports, driver's licenses, offer letters with Claude vision
→ Auto-creates HRIS records, IAM accounts, onboarding checklists
→ Sends I-9 / W-4 to eSignature, enrolls benefits, orders background checks
→ Tracks compliance deadlines (I-9 at hire+3 days, state new-hire reporting at hire+20)
→ Flags REVIEW_REQUIRED for HR. Human in the loop, always.
But the real lesson isn't in any of that.
Background checks come back as "clear," "consider," or "fail." A naive agent auto-rejects on "consider." That's where compliance breaks. EEOC requires individualized human assessment, not algorithmic rejection.
The fix wasn't a bigger system prompt. It was ONE line in the tool description:
"If result is 'consider,' alert HR for manual review. Never auto-block."
Production agents aren't built in the system prompt.
They're built in the tool descriptions.
Stack: Python
· Claude Agent SDK (sonnet-4-6 + vision)
· AgentAPIsHub
· httpx
· Pydantic
· Rich 9 modules
· 26 tools
· 1,200 lines.
🎥 Full walkthrough + downloadable source:
Question for builders here. What's one tool description you've written that quietly saved you from a production bug? Drop it below.