An IT agent that's allowed to say no

The brief

Build agents that act: agents that change real systems, not ones that only chat about them.

The problem I picked

Employee lifecycle IT. Onboarding is days of tickets: an inbox, repo access, cloud keys. Offboarding is worse, because whatever nobody remembers to revoke becomes a zombie credential.

What the agent actually touches

Through a FastMCP server, with TrueForge as the reasoning engine and the policy written as a SKILL.md:

• Zoho Directory: creates and suspends real @vansshagarrwal.in inboxes.
• GitHub: sends collaborator invites and manages org permissions.
• AWS IAM: issues and invalidates scoped developer keys.
• MongoDB Atlas: keeps an audit record for every lifecycle change.
• Zoho SMTP: delivers credentials out of band, never in the chat.

Where it stops

This was the part I cared about most. An agent with IAM access is only as good as its refusals:

• Zero-trust repos: requests for sensitive repositories are denied outright.
• Human-in-the-loop: offboarding is irreversible, so it waits for a manager’s explicit confirmation.
• Disambiguation guard: if two active employees share a name, it halts instead of guessing.
• No secrets in chat: AWS secret keys never appear in the conversation or the dashboard.

Real vs mocked

The README says it plainly: inboxes, GitHub invites, IAM keys, Atlas writes and emails are real. The billing_core 403 trap and the Daytona sandbox probe are simulated. In a demo, being explicit about this is worth more than pretending everything is live.