OnboardOps

Inspiration

New employee onboarding is usually scattered across emails, spreadsheets, and Slack messages - HR paperwork, IT provisioning, scheduling, and policy questions all tracked manually. Both new hires and managers lose time chasing status updates instead of getting real work done. We wanted to see if an AI agent could own that entire coordination burden instead of a human doing it by hand.

What it does

OnboardOps is an AI agent that automates onboarding end-to-end. It generates a personalized, role-specific onboarding checklist; answers common HR policy questions (leave, benefits, reimbursement); builds a suggested first-week meeting schedule; and logs every status update to a shared Google Sheet tracker giving managers full visibility without manual follow-up. A manager can just talk to it in plain language across a conversation ("she's starting Monday, generate her checklist and schedule, and tell me the leave policy") and it chains the right tools together on its own.

How we built it

A single Strands Agent orchestrator, running on Amazon Bedrock (Claude), routes requests to four specialized tools: a checklist generator using Pydantic-structured output grounded in a maintained Indian HR/labour-law compliance reference; a fuzzy-matching HR Q&A tool over a policy knowledge base; a scheduling tool that builds a first-week agenda from a start date; and a Google Sheets tracker with four functions (log, retrieve, update, view-all status). Each teammate owned one tool, tested it standalone, then handed it off for integration into the shared orchestrator.

Challenges we ran into

Getting the agent to use its tools only when explicitly asked was harder than expected early versions would sometimes log tasks to the tracker on their own initiative after just generating a checklist, and occasionally lost track of a name or detail mentioned earlier in a multi-turn conversation. We fixed this with an explicit system prompt constraining tool-use behavior, and verified it with realistic multi-turn conversation tests simulating an actual manager. We also hit a checklist→tracker format mismatch: checklist items are long, detailed sentences (100-360+ characters), but the tracker needs short task labels solved with a helper that derives a short, deterministic label from each item without an extra LLM call.

Accomplishments that we're proud of

All four tools work independently and chained together through a single agent, verified end-to-end with real prompts routing correctly and writing live rows to the shared tracker. The checklist tool grounds statutory items in real compliance references (named forms, current thresholds) instead of relying on the model's own recall and it genuinely adapts its output by role and seniority, not just department.

What we learned

Constraining an agent's tool-calling behavior with explicit instructions matters as much as building the tools themselves, a capable agent with loose instructions will do more than you asked. We also learned the value of catching integration issues (like the checklist-to-tracker format mismatch) early with real, non-mocked test data instead of hardcoded happy-path examples.

What's next for OnboardOps

Adding task IDs to the tracker schema so tasks can be reliably referenced and updated across regenerated checklists, deploying via Amazon Bedrock AgentCore, and expanding the HR knowledge base with more nuanced, context-aware matching for vague policy questions.