Ship It: The Pre-Launch Checklist

Your agent works in every test you’ve thrown at it, the demo goes smoothly, and everyone’s ready to call it done. Sit with this for a second: that exact feeling — “it works, ship it” — is precisely how every agent that later caused a production incident felt the day before launch too. Confidence isn’t evidence. This checklist is.

The five checks

1 · Scope — back to Lesson 2

Reread your one-sentence job description out loud. Does the agent, as it exists today, actually do that — no more? Grown two tools past it? Update the sentence deliberately; don’t let it drift unnoticed.

2 · Write-safety — back to Lesson 4

Every write: correct create sequence, mandatory fields accounted for? Update fetches first, so the ETag is always captured? Any delete capability that should actually be a block?

3 · Testing — back to AI Foundations Lesson 12

A real eval set — golden questions, expected behaviour, a must-refuse row? Run more than once, with anything that flipped noted? “We tried it a few times and it seemed fine” is a demo you got lucky with, not an eval set.

4 · Governance — back to AI Foundations Lesson 13

Which actions need human approval, which are safely autonomous? Does authorization strictly inherit from the acting user — no shared service account with broader rights? What exactly gets logged? Have you actually tried one prompt-injection attempt, or only ever tested clean inputs?

5 · Clean Core — the foundation under all four

Does every tool call a released API, zero exceptions, zero internal function modules “just this once”? This one item isn’t optional. Everything above is risk management. This is the foundation it’s built on.

Try it yourself — the real capstone

Run your own agent through all five, honestly, in writing — not “confident it’s fine,” an actual answer for each, including the ones you’d rather skip. Anywhere you can’t answer with real confidence is exactly where the next hour of work goes, before launch, not after.

An agent that passes this checklist honestly is one you can put your name behind in a client’s security review — not because it’s perfect, but because you can defend every decision, one by one.

You’ve finished Build SAP AI Agents

From a bare BTP subaccount to a scoped, tested, governed agent running in Joule — that’s the full arc. Combined with SAP AI Foundations, you now hold both the conceptual grounding and the hands-on discipline to build something a client can trust in production, not just admire in a demo.

Where to go next:

• Full working code, end to end: Business Partner and Sales Order tutorials.
• Make the foundation itself safe to build on: the S/4HANA Migration & Clean Core journey.
• Check what’s blocking your own landscape: a free readiness scan.

Key takeaways

• Scope, write-safety, testing, governance, Clean Core — five checks, none optional, all traced to specific lessons.
• “It worked in every test” is a feeling. A real eval set with a must-refuse row is evidence.
• Clean Core isn’t one check among five equals — it’s the foundation the other four are built on.
• Passing this checklist honestly means you can defend the agent in detail — that’s the actual bar for production.