The most useful thing this AI agent did was tell me it wasn't done yet

We have been writing about MCP's potential since its earliest days. So when Workday shipped support for third-party agents, we finally got to move from writing to wiring.

We connected Claude Code into a live Workday GMS as a registered agent and had it change a worker's business title end to end:

The moment that matters

The agent looks up the worker, drafts the exact change, and then stops:

It will not execute without an explicit yes, and it will not claim success until it re-reads Workday and sees the change with its own eyes.

The live session: read-only lookup, proposal with the confirm token held, explicit approval, then honest verification. The write landed as a Title Change event, and the agent reports it as submitted but not confirmed because the worker record still shows the old title, pending business process approval.

Four things make that possible

• Approval sealed to the exact change: a cryptographic token binds the approval to one specific change. Approve one thing, and only that thing can be written.
• Verification before "done": a hook blocks the agent from claiming success without re-checking Workday. A chat transcript is never evidence.
• Dual authorization on every call: Workday checks two things on every single request: is this user allowed to use this skill, and is this user allowed to perform the underlying operation. The agent runs as a registered identity in the Agent System of Record, every request passes through one Agent Gateway, and it inherits nothing beyond the approving human's own permissions.
• Full accountability: the Workday audit trail records every action the agent took and the named human it acted for.

Your business process still governs. Our test change is sitting in an approval queue right now, exactly where it belongs.

What's next

How the token binding works, why the write sat in that approval queue, and the full Workday configuration behind it: that's a longer story. The detailed Claude-Workday build article is coming soon here on Perspectives. Follow along if you're putting agents anywhere near a system of record.