Who is accountable for your APEX AI Agent?

I recently read a LinkedIn post from an old friend that got me thinking about AI and accountability. More specifically, who is accountable for what your APEX AI agents do?

TL;DR: As an APEX developer, you are responsible for the tools you expose and the actions those tools allow.

APEX passes the user's request and available tool definitions to the LLM. The LLM can then ask APEX to run tools to complete the request. You decide which tools are available, what they do, and what checks they enforce.

If the LLM asks to call the delete_employee tool for every employee in your database and your tool allows it to happen, you are responsible for the missing safeguards.

APEX 26.1 provides tool authorization schemes and, for supported front-end components, user confirmation before sensitive tools run. Use them.

Your server-side code must still check every requested action:

• Who is the user, based on the authenticated session?

• Can that user access the record and perform this action?

• Are the supplied parameters valid?

• Do business rules allow the action?

These checks can live in shared PL/SQL packages. They need to run for every tool call.

But permission alone is not enough. A user might have permission to delete every employee without ever intending to do so. Destructive tools need limits across the whole operation, plus confirmation that makes the affected records clear. Accepting thousands of individually valid calls can still produce a bad result.

The LLM will sometimes misunderstand a request. Your tools determine how much damage that misunderstanding can do.