Skip to course content
Free agentic AI fundamentals course

Introduction to Agentic AI and Workflow Automation

Unit 03.02: Validating what the agent asked for

Arguments produced by an agent are untrusted input that happens to have arrived from inside your own system.

Shape checks and policy checks

Some checks are about type and format. Others encode a business limit, and those are the ones worth being deliberate about.

The example below validates three refund requests.

POLICY LIMITS
  maximum refund     500.00
  currencies         USD, EUR
  account format     ACC- followed by digits

PROPOSED ACTION                                    VERDICT
account ACC-1187, 240.00 USD                       ALLOW
account ACC-1187, 9000.00 USD                      BLOCK  amount exceeds 500 limit
account 1187, 240.00 BTC                           BLOCK  currency not permitted
                                                          account format wrong

  The 9,000 request is a well-formed number and a well-formed account. It
  fails on policy, not on shape -- which is why the limit is a written rule
  and not a type.

The 9,000 request passes every type check - it is a number, it is positive, the account is well formed - and fails the policy one. That limit is not a validation detail; it is a decision about how much this automation may move, and it belongs in a written rule where an auditor can read it and a check can enforce it.

The same limit written into an instruction is a request to a component that has already been handed the amount.

The mistake this prevents

The mistake is validating types and stopping. A schema will accept a refund of ninety thousand because ninety thousand is a number. Write the ceiling separately, as data, near the top of the file.

Takeaway

Validate shape and policy as separate layers. Policy limits belong in a written rule, where they can be audited and checked.